Continuously tracking of moving object by a combination of ultra-high frequency radio-frequency identification and laser range finder

Abstract

Due to the unique and contactless way of identification, radio-frequency identification is becoming an emerging technology for objects tracking. As radio-frequency identification does not provide any distance or bearing information, positioning using radio-frequency identification sensor itself is challenging. Two-dimensional laser range finders can provide the distance to the objects but require complicated recognition algorithms to acquire the identity of object. This article proposes an innovative method to track the locations of dynamic objects by combining radio-frequency identification and laser ranging information. We first segment the laser ranging data into clusters using density-based spatial clustering of applications with noise (DBSCAN). Velocity matching–based approach is used to track the location of object when the object is in the radio-frequency identification reading range. Since the radio-frequency identification reading range is smaller than a two-dimensional laser range finder, velocity matching–based approach fails to track location of the object when the radio-frequency identification reading is not available. In this case, our approach uses the clustering results from density-based spatial clustering of applications with noise to continuously track the moving object. Finally, we verified our approach on a Scitos robot in an indoor environment, and our results show that the proposed approach reaches a positioning accuracy of 0.43 m, which is an improvement of 67.6% and 84.1% as compared to laser-based and velocity matching–based approaches, respectively.

Keywords

Ultra-high frequency radio-frequency identification density-based spatial clustering of applications with noise clustering velocity matching continuous tracking particle filter

Introduction

With the development of IoT (Internet of Things), there is an increasing demand for location-based services (LBSs). Global positioning system (GPS) is superb at positioning in outdoor environments, but it becomes useless in indoor environments due to the occlusions of the buildings.¹ Indoor localization draws more and more attention to the researchers in recent years. For example, Yasir et al.² introduced an indoor positioning system based on visible light and accelerometer measurements. Zhuang et al.³ proposed an indoor positioning platform that integrated different wireless access points for indoor positioning. As one of the key components of the IoT, radio-frequency identification (RFID) provides a cost-effective solution for the identification of object.^4,5 The tags are small and does not need any batteries, which make them suitable for the tracking of the assets in many industrial environments. Especially, the long-range passive ultra-high frequency (UHF) RFID technology can provide a reading range up to 10 m. As compared with vision and laser technologies, RFID has the advantages of non-contact and non-line-of-sight, which have the potential for tracking objects in many commercial environments, for example retail and logistics.⁶

Based on the information provided by the reader (signal strength or phase), different techniques are used for the positioning purposes, for example, received signal strength (RSS),^7,8 angle of arrival (AOA),⁹ time of arrival (TOA),¹⁰ and time difference of arrival (TDOA).¹¹ For example, Liu and colleagues^12,13 proposed a LANDMARC system that achieves the localization of object by comparing the signal strength from multiple reference tags. They used a weighting algorithm to calculate the position of an unknown tag. Instead of deploying multiple readers in the environment, the system used reference tags to reduce the overall system cost. A number of researchers exploit the signal strength for the positioning of RFID tags.^14–16 These approaches require modeling of the signal propagation, which is usually very hard in practice and is easily influenced by many environmental factors. The use of the AOA and the TOA requires high hardware costs and therefore has low practicability. In addition, one can also use RFID phase information for positioning. But this method has the problem of phase ambiguity, which leads to a very limited positioning accuracy. Some researchers proposed multi-frequency carrier¹⁷ and linear-frequency modulation¹⁸ methods to resolve phase ambiguity, but these methods have higher requirements for hardware.

Although RFID can be used to identify an object, RFID does not provide the distance or the bearing information to an object. The laser range finder can provide accurate distance information according to the laser beam. Therefore, many researchers use laser range finder to locate and track objects. For example, an event-based algorithm is proposed for the localization of a mobile robot by using a fixed laser range finder.¹⁹ Wang et al.²⁰ used a single Lidar to achieve indoor positioning of a mobile robot based on weighted parallel iterative closed point (WP-ICP) with interpolation. The WP-ICP algorithm results in faster convergence for pose estimation, but it is easily affected by static or dynamic unknown objects in the environment. Although a precise distance information can be obtained from a laser range finder, it needs complex algorithms to recognize the object from a large amount of scanning points. In order to do this, a prior knowledge of the object has to be known in advance, for example, shape of the object.

As different sensors might have different drawbacks, which can be compensated for each other, many researches use sensor fusion techniques to achieve a robust and precise indoor positioning system. Zhang et al.²¹ fused RFID and inertial navigation system (INS) based on extended Kalman filter (EKF) to achieve object localization in a multi-story building. An adaptive sensor fusion method which automatically compensates for bias between a laser range finder and video camera was presented.²² Chen et al.²³ proposed an environment knowledge-based indoor positioning system by integrating Lidar, odometer, and light sensor with an EKF. The integration of RFID and laser ranging information for positioning also attracts many researchers’ attentions. From a perspective of sensor fusion, the laser range finder provides distance to an object, while the RFID directly identifies the object. Therefore, RFID and laser can complement each other’s shortcomings. However, in practical applications, the reading ranges of RFID and laser are different. The range of laser sensor is usually larger than the RFID reading range. In order to continuously track dynamic object when the object is out of the RFID reading range, we propose a dynamic object localization method based on the fusion of RFID and laser ranging information. When the object is in the range of RFID sensor, we use velocity matching to fuse RFID and laser information and then apply a particle filter to track the object. This idea is also used in many fields, for example, dynamic target positioning²⁴ and alignment of different sensor measurements.²⁵ To segment the raw laser ranging measurements into clusters, we use the density-based spatial clustering of applications with noise (DBSCAN) clustering algorithm.²⁶ The algorithm can quickly find clusters of arbitrary shapes and deal with noise points,²⁷ which is widely used in the fields of image processing and data mining. Otherwise, we track the object based on clustering results and the previous estimation. Our approach expands the tracking range as compared to our previous research and ensures a continuous tracking of an object when the RFID information is not available. The proposed system can be applied to many areas such as human-robot interaction and object tracking.

The remainder of the article is organized as follows. The “Related work” section reviews the existing work RFID-based approach, dynamic target-tracking approach, fusion with laser ranging information, and fusion with other sensor. The “Overview” section briefly describes our positioning system. The “Laser clustering based on DBSCAN” section presents the working principle of the DBSCAN clustering algorithm and calculates the laser-based velocity based on the clustering results. The “Velocity matching” section describes the calculation of phase-based velocity and calculates the similarity between the laser-based velocity and the phase-based velocity to achieve velocity matching. “Tracking the dynamic object based on the particle filter” section presents the object tracking using a particle filter. “Experiment” section shows the experimental setup and results. Finally, “Conclusion” section presents the conclusion and the future work of this article.

Related work

This section gives a thorough overview of the work related to the tracking using RFID techniques. We divide the related work into the following three categories: RFID-based approach, fusion with laser ranging information, and the fusion with other sensors. Meanwhile, we discuss their continuous tracking capabilities and their advantages and disadvantages.

RFID-based approach

Many RFID readers report the signal strength of the detected tags, which can be used to infer the location of the tag. Two typical systems using this technique are VIrtual Reference Elimination (VIRE)²⁸ and SpotON.²⁹ AOA is also exploited for the positioning of RFID by several researchers. Zhou et al.³⁰ estimated the AOA by phase difference and showed the feasibility to use AOA for the localization. The phase is easily affected by environmental factors as well as the radiation patterns of reader and tag antennas. TOA shows very good accuracy for positioning, but it is difficult to achieve time synchronization. To overcome this problem, time difference of arrival (TDOA) has been employed, but the measurement is susceptible to interference from non-line-of-sight factors. A location-tracking system is proposed based on measuring the phases from multiple spatially distributed antennas at a single carrier frequency.³¹ Kim et al.³² proposed a method to fuse TDOA and AOA for positioning using dual indirect Kalman filter (DIKF). A fine-grained positioning system was presented using the phase information from the COTS RFID products.³³

Dynamic target-tracking approach

Dynamic target tracking is widely used in intelligent monitoring, human–robot interaction, virtual reality, and robot navigation. For example, Zou et al.³⁴ combined image and point cloud data to construct map of an environment and track the position of moving targets in the view. To improve the tracking accuracy of dynamic targets, they used the three-dimensional (3D) motion model and the shape of the target to deal with obstacle occlusion problem. Lopes et al.³⁵ proposed a dynamic target-tracking system using blur images. Li et al.³⁶ proposed an adaptive distributed sampling online co-search method to achieve dynamic target tracking. They formed the posterior probabilities of an unpredictable moving target using only local information and constructed a piecewise function to describe the likelihood of measurements. Xiao et al.³⁷ designed an algorithm to track arbitrary objects in the video sequences based on adaptive clustered decision tree.

Fusion with laser ranging information

As laser range finders provide very precise distance measurement to the surroundings, some researchers exploited the possibilities to integrate RFID and laser ranging information for positioning. For example, Lin et al.³⁸ proposed an indoor mapping approach that combines active RFID and laser technology. They initialized the robot’s global configuration by recognizing active RFID tags in the indoor environment and used the laser range finder to build the metric map of an environment. Gong et al.³⁹ presented an object-positioning algorithm based on RFID and laser range finder. They used RFID to estimate the rough existing area of the object and obtained a precise position of the object using a laser feature–detection algorithm. He et al.⁴⁰ used the RFID phase difference and the laser to calculate the object velocity, and then performed velocity matching. Then, based on the matching result, they located the dynamic target according to a particle filter. In our previous work,⁴¹ we proposed a way for the localization of moving object using a particle filter by incorporating RFID phase and laser-based clustering from two-dimensional (2D) laser range data. We achieved the tracking of the object by a comparison of the velocity between the RFID phase velocity and laser-based velocity. However, this method requires the object to be in the range of the RFID reader. In this article, we show that we can achieve object tracking even if it is out of the RFID detection range.

Fusion with other sensors

Zampella et al.⁴² presented an indoor positioning system that fused foot-mounted inertial measurements and radio frequency (RF) measurements based on a particle filter. The RF system consisted of an RFID tag in the waist and an ultra wide band (UWB) receiver in the chest. In order to overcome the problem of insufficient coverage of RFID signals, some researchers proposed to combine RFID and INS through Kalman filtering to achieve continuous tracking of pedestrian.^43,44 A positioning system that fuses RFID and sonar sensors was presented by Choi et al.⁴⁵ Xie et al.⁴⁶ proposed an object-recognition method based on RFID and depth camera. Liu et al.⁴⁷ proposed a method for quickly tracking dynamic targets based on the RSS measurements from a pair of RFID antennas. Duan et al.⁴⁸ proposed Tagvision that achieved the tracking of fine-grained objects by combining computer vision and RFID. This method combined the location information provided by the vision system with the phase data from the RFID to achieve a good accuracy.

Table 1 shows a comparison of different approaches. The existing approaches either requires a prior knowledge about the propagation of the radio signal, or fails to track the moving object when it is out of the RFID reading range. The proposed approach does not need any modeling of the signal propagation and allows the continuous tracking of an object when the object is not in the range of the RFID reader.

Table 1.

Comparison of different approaches.

Researchers	Approach	Accuracy (m)	Remarks	Continuoustracking
Zhao et al.²⁸	RSS of the virtual reference tags	0.14–0.47 m	Need reference tag array	No
Zhou et al.³⁰	AOA by phase difference	1.7 m	Vulnerable to the environment	No
Kim et al.³²	Fusing TDOA and AOA based on Kalman ﬁlter	0–350 m	Large error and high hardware requirements	Yes
Liu et al.³³	Phase from COTS RFID	0.13 m	Vulnerable to the environment and small range	No
Lopes et al.³⁵	Image blurring	3.3 m	Need multiple features not efﬁcient	Yes
Lin et al.³⁸	Combine active RFID and laser	0.20 m	High cost and need to build a model	Yes
Gong et al.³⁹	RFID and laser feature detection	1 m	Need to build model	Yes
He et al.⁴⁰	RFID phase and laser	0.3 m	Positioning range is small	No
Previous approach⁴¹	Fusing RFID and laser based on PF	0.25 m	Positioning range is small	No
Zampella et al.⁴²	RFID and UWB	≤2 m	High hardware requirements	Yes
Retscher and Fu⁴⁴	Combined RFID with INS through KF	1–2 m	Need to build a model	Yes
Xie et al.⁴⁶	Object matching based on RFID and depth camera	91% match ratio	Power consumption andnot portable	Yes
Liu et al.⁴⁷	RFID combined with particle ﬁlter and VFH	0.27–0.38 m	Need to build a model	Yes
Our approach	Continuously tracking by RFID and laser	0.43 m	Model free and large coverage	Yes

RSS: received signal strength; AOA: angle of arrival; TDOA: time difference of arrival; RFID: radio-frequency identification; UWB: ultra wide band; INS: inertial navigation system; KF: Kalman filter.

Overview

This article introduces a positioning method that fuses RFID and 2D laser range finder based on the particle filtering for the tracking of a dynamic object. The proposed approach allows the continuous tracking of an object when the object is out of the RFID reading range. As shown in Figure 1, the entire positioning system can be divided into three parts. The first part uses RFID reader and 2D laser range finder to perceive the measurements from the environment. The second part processes the measurements collected from RFID and laser range finder. The third part uses a particle filter to achieve tracking and localization of a dynamic object. The mathematical symbols and their meanings used in this article are listed in Table 2.

Figure 1.

System overview.

Table 2.

Mathematical symbols and their meanings.

Symbol	Meaning
$ϵ$	Radius of cluster
$ζ$	The minimum number of samples included in the cluster
$N_{ϵ}$	Neighbors within radius of $ϵ$
$C_{t}$	A set of clusters at time $t$
$C_{t}^{(i)}$	The ith cluster at time $t$
$k_{t}$	The number of clusters at time $t$
${\bar{P}}_{t}^{(j)}$	The center of each cluster at time $t$
$v_{t}^{(j)}$	The laser-based velocity of cluster $j$ at time $t$
$θ_{t}$	Phase of RF signal at time $t$
$v_{t}^{r}$	Phase-based velocity of RFID tag at time $t$
$K$	The number of the best matching clusters
$X_{t}$	The object position at time $t$
$a_{1 : t}$	RFID measurements at $1 : t$
$l_{1 : t}$	laser ranging measurements at $1 : t$
$η_{t}$	Normalization factor at time $t$
$u_{t}$	Motion information of object at time $t$
$N$	Number of particles
$X_{t}^{[n]}$	The coordinates of particle $n$ at time $t$
$ω_{t}^{[n]}$	The weight of particle $n$ at time $t$
$σ_{d}$	The translational coefficient in the distance function
$d_{\max}$	The maximum distance of object move at $Δ t$
$Δ d_{t}^{(j)}$	The distance between $C_{t}^{(j)}$ and the estimated position

RFID: radio-frequency identification.

In the first part, we collect the measurements through RFID reader and 2D laser range finder. The RFID reader controls the antenna to scan the tagged dynamic objects in the environment, and the 2D laser sensor is used to measure the distance and angle from the surrounding obstacles in an environment. Then the information detected by the two sensors is stored the database. The second part processes the measurements collected from the previous part. In this part, we process the acquired data separately based on whether RFID information is available. When the moving object is in the range of RFID sensor, we combine RFID and laser information for the tracking. First, we calculate the phase-based velocity of the moving object by the RFID phase difference between two successive timestamps. The laser ranging data are segmented into different clusters using DBSCAN.⁴⁹ The two closest clusters at the adjacent time are regarded as the same obstacle, and the distance between the two clusters is used to estimate the laser-based velocity. Second, we perform velocity matching to compute the similarity between phase-based velocity and laser-based velocity. Based on the matching results, we choose the best $K$ match clusters to update the weights of the particles. When the object is out of the RFID reading range, we use DBSCAN clustering results to achieve the continuous tracking of the object. To be precise, we select the cluster closest to the previous estimated position and consider this cluster as the object to be tracked. In the third part, we fuse the data of the second part to achieve the dynamic object tracking using a particle filter. The particle filter performs initialization and iteratively performs prediction, update, and resampling steps which will be detailed in the next section of this article.

Laser clustering based on DBSCAN

DBSCAN clustering of laser ranging data

The 2D laser range finder records distance information to the objects in the scanning environment. To facilitate subsequent clustering, we assume $(x_{i}, y_{i})$ is the coordinate of the $i th$ laser point at time $t$ . Figure 2(a) shows an example of the raw laser data, where the black circle denotes the moving object.

Figure 2.

Example of the raw laser data at a time.

In order to better determine the position of the object, we need to cluster the laser ranging data. As can be seen from Figure 2, the laser points around the obstacle are more concentrated, and we can cluster the laser ranging points based on theirs densities. DBSCAN is a density-based clustering algorithm, which can cluster dense points of arbitrary shape and remove the abnormal points. Therefore, this article chooses DBSCAN algorithm to cluster the two-dimensional laser ranging data. In DBSCAN clustering, for each point in a cluster, the number of its $ϵ$ neighbors $N_{ϵ}$ must exceed a given threshold $ζ$ . The un-clustered points in the ranging measurements are considered as noise.

For DBSCAN clustering, we first assign an initial mark −1 to all laser points. From this mark, we can determine whether the laser point has been processed. Next, we divide the marked laser points into core points, border points, and noise points. More specifically, for a given point $i$ (i.e. the $i th$ laser beam returned by a laser range finder), we find its $ϵ$ neighbors $N_{ϵ} (i)$ whose Euclidean distances to point $i$ are smaller than $ϵ$ . If $N_{ϵ} (i)$ is higher than a given value $ζ$ , $i$ is regarded as the core point. If $N_{ϵ} (i)$ is smaller than $ζ$ but within neighborhoods of a core point, $i$ is considered as the border point. Otherwise, we treat $i$ as the noise point. The flow chart of DBSCAN clustering is shown in Figure 3.

Figure 3.

DBSCAN flow chart.

We perform clustering after finding the core points and border points. If a point is the core point, we form a cluster by expanding its $ϵ$ neighbors. For each border point, it will be assigned to a nearby cluster if the border point is an $ϵ$ neighbor of that cluster, otherwise it is regarded as noise. Finally, we obtain a set of clusters after performing DBSCAN on the laser ranging measurements in an indoor environment: $C_{t} = {C_{t}^{(1)}, \dots, C_{t}^{(k_{t})}}$ , where $k_{t}$ is the number of clusters at time $t$ . The center of each cluster is denoted as ${\bar{P}}_{t}^{(j)} = ({\bar{x}}_{t}^{(j)}, {\bar{y}}_{t}^{(j)})$ . Figure 2(b) shows the clustering results using DBSCAN. The performance of DBSCAN is impacted by two parameters $ϵ$ and $ζ$ , which will be analyzed in our experimental results.

Estimate moving velocity of a cluster

For each cluster $C_{t}^{(j)}$ , we find the nearest cluster $C_{t - 1}^{(\hat{j})}$ at the previous timestamp and consider these two clusters as the same moving object at two different times. $\hat{j}$ is given by

\hat{j} = \arg min_{i} \sqrt{{(x_{t}^{(j)} - x_{t - 1}^{(i)})}^{2} + {(y_{t}^{(j)} - y_{t - 1}^{(i)})}^{2}}

(1)

where $1 \leq i \leq k_{t - 1}$ . After DBSCAN, the radial velocity $v_{t}^{(j)}$ of cluster $j$ at time $t$ is computed as

v_{t}^{(j)} = \frac{\sqrt{{(x_{t}^{(j)})}^{2} + {(y_{t}^{(j)})}^{2}} - \sqrt{{(x_{t - 1}^{(\hat{j})})}^{2} + {(y_{t - 1}^{(\hat{j})})}^{2}}}{Δ t}

(2)

The velocities of the clusters will be used for velocity matching in Section “RFID and laser velocity matching.”

Velocity matching

RFID phase velocity estimation

This article uses RFID phase information to estimate the velocity of the object, which is referred to as phase velocity in this article. The phase information of the RFID is a periodic function (with a period of $2 π$ ) of the distance between the tag and the antenna^16,50

θ_{t} = 2 π \cdot \frac{2 d_{t}}{λ} \mod (2 π)

(3)

where $θ_{t}$ represents the phase of the signal at time $t$ , $λ$ is the wavelength of the radio signal, and $d_{t}$ represents the distance from the RFID tag to the antenna at time $t$ . Equation (3) explains the relationship between the phase the distance. It is well known that the received signal strength decreases with the distance, but we use the relative phase change to measure the radial velocity of a moving object, which is shown to be more robust against the distance. We can use the phase difference at two timestamps to estimate the moving velocity $v_{t}^{r}$ of the object

v_{t}^{r} = {\begin{matrix} \frac{c}{4 π f Δ t} \cdot (θ_{t} - θ_{t - 1}), & | θ_{t} - θ_{t - 1} | < π \\ \frac{c}{4 π f Δ t} \cdot (θ_{t} - θ_{t - 1} - 2 π), & θ_{t} - θ_{t - 1} \geq π \\ \frac{c}{4 π f Δ t} \cdot (θ_{t} - θ_{t - 1} + 2 π), & θ_{t} - θ_{t - 1} \leq - π \end{matrix}

(4)

where $c$ is the velocity of light $(3 \times 10^{8} m / s)$ , $f$ represents the frequency of the radio signal. To eliminate the phase ambiguity, $Δ d_{t}$ is less than half of the wavelength (i.e. $Δ d_{t} < λ / 2$ ).

RFID and laser velocity matching

At a specific time, there may be many moving or static objects (i.e. clusters) in the environment. In the simplest case, the cluster with the closest velocity to the phase velocity is considered to be the moving object we want to track. Therefore, for a given cluster, we perform velocity matching, that is, compare its velocity against the phase-based velocity by a similarity measures

sim (v_{t}^{(j)}, v_{t}^{r}) = 1 - \frac{| v_{t}^{(j)} - v_{t}^{r} |}{| v_{t}^{(j)} | + | v_{t}^{r} |}

(5)

where $sim (v_{t}^{(j)}, v_{t}^{r})$ denotes the similarity between two velocities $v_{t}^{(j)}$ and $v_{t}^{r}$ . To improve the robustness of the system, we choose $K$ clusters with the highest similarity values and use the corresponding clusters $C_{t}^{m (1)}, C_{t}^{m (2)}, \dots, C_{t}^{m (k)}$ to update the particle filter.

Tracking the dynamic object based on the particle filter

Our goal is to estimate the location $X_{t}$ of a moving object at time $t$ , given a series of RFID and laser ranging measurements. Let $a_{1 : t}$ and $l_{1 : t}$ be the RFID and laser ranging measurements at $1 : t$ in an indoor scene, respectively. $u_{1 : t}$ denotes the motion information of the object in an indoor scene. The objective is to estimate the posterior probability density distribution $p (X_{t} | l_{1 : t}, a_{1 : t}, u_{1 : t})$ , given a sequence of motion measurements $u_{1 : t}$ and sensor measurements $a_{1 : t}$ and $l_{1 : t}$ . According to the Bayesian theory, $p (X_{t} | l_{1 : t}, a_{1 : t}, u_{1 : t})$ is decomposed as

\begin{matrix} p (X_{t} | l_{1 : t}, a_{1 : t}, u_{1 : t}) = η_{t} \cdot p (X_{t} | X_{t - 1}, u_{t}) \cdot p (a_{t} | X_{t}, l_{t}) \\ \cdot p (X_{t - 1} | l_{1 : t - 1}, a_{1 : t - 1}, u_{1 : t - 1}) \end{matrix}

(6)

where $η_{t}$ is the normalization factor to make sure the sum of probability equals to one. $p (X_{t} | X_{t - 1}, u_{t})$ represents the motion model of the target in the indoor environment, which is used to predict the current location of the object based on the previous location and motion information $u_{t}$ . $p (a_{t} | X_{t}, l_{t})$ is the observation model in the indoor scene, which represents the likelihood of receiving the RFID measurement $a_{t}$ , given the state $X_{t}$ and the laser ranging information $l_{t}$ . $p (X_{t - 1} | l_{1 : t - 1}, a_{1 : t - 1}, u_{1 : t - 1})$ is the posterior probability density function at $t - 1$ .

In the literature, many implementations of Bayesian filter have been proposed. The particle filter demonstrates good capability in nonlinear and non-Gaussian systems, and has been successfully used in many fields, for example object tracking, navigation, and positioning. Therefore, this article chooses the particle filtering as the implementation. In the particle filtering, the probability density function is represented by a set of particles, that is, $X_{t} = {X_{t}^{[n]}, ω_{t}^{[n]}}_{n = 1}^{N}$ , where $N$ is the number of particles randomly placed in an indoor environment, $X_{t}^{[n]} = (x_{t}^{[n]}, y_{t}^{[n]})$ is the 2D location hypothesis and $ω_{t}^{[n]}$ is the weight. The location of the object is estimated by a weighted mean among all particles. The particle filtering will perform prediction, update, and resampling iteratively based on the arrival of the measurements.

Prediction

This part predicts the particle location $X_{t}$ according to the motion model and the previous location $X_{t - 1}$ . Since we do not have any knowledge about the moving direction of the object, the prediction is performed with the following

{\begin{matrix} x_{t}^{[n]} = x_{t - 1}^{[n]} + N (0, σ_{t}) \\ y_{t}^{[n]} = y_{t - 1}^{[n]} + N (0, σ_{t}) \end{matrix}

(7)

where $N (0, σ_{t})$ is the Gaussian noise with 0 mean and standard deviation of $σ_{t}$ . In order to better predict the object location, we determine $σ_{t}$ based on the moving speed of the object. When the object is in the RFID reading range, we find the laser cluster $j^{*}$ , whose velocity best matches the RFID phase velocity $v_{t}^{r}$ and use $Δ t \cdot v_{t}^{(j^{*})}$ as a representation of $σ_{t}$ , where $Δ t$ is the time difference between two timestamps. Otherwise, when the tag is out of the RFID reading range, we set $σ_{t} = Δ t$ by assuming that the maximum moving speed of a person does not exceed 1.0 m/s.

Update

The update stage computes the weight of each particle based on the sensor measurements. Based on the availability of the RFID information, we need to update the particle weights according to the following two cases

1. Update weights using $K$ best match when RFID information is available.

For the case that the RFID signal is available, the weights are updated based on the $K$ best match clusters after the velocity matching. The weight $ω_{t}^{[n]}$ of a particle $X_{t}^{[n]}$ is updated according to

ω_{t}^{[n]} = η_{t} \cdot ω_{t - 1}^{[n]} \cdot \sum_{i = 1}^{K} sim (v_{t}^{m (i)}, v_{t}^{r}) \cdot \exp (- \frac{d^{2} (X_{t}^{[n]}, C_{t}^{m (i)})}{2})

(8)

d^{2} (X_{t}^{[n]}, C_{t}^{m (i)}) = \frac{{(x_{t}^{[n]} - {\bar{x}}_{t}^{m (i)})}^{2}}{σ_{d}^{2}} + \frac{{(y_{t}^{[n]} - {\bar{y}}_{t}^{m (i)})}^{2}}{σ_{d}^{2}}

(9)

where $σ_{d}$ is the translational coefficient in the distance function, which affects the weight distribution of particle filtering.

2. Update weights using laser cluster when the RFID information is not available.

When the RFID information is not available, we use the laser ranging data alone for the updating of the particle filter. This ensures the continuous tracking of the object when the object is out of the RFID reading range. In reality, the distance that an object moves within a time interval should not exceed a certain range $d_{\max}$ . Therefore, we calculate the distance $Δ d_{t}^{(j)}$ between cluster $C_{t}^{(j)}$ and the estimated position of the object. If $Δ d_{t}^{(j)} \leq d_{\max}$ , the corresponding cluster $C_{t}^{(j)}$ is regarded as the moving object. There may be more than one cluster satisfying $Δ d_{t}^{(j)} \leq d_{\max}$ . We choose a cluster $\hat{l}$ that is closest to $(x_{t - 1}, y_{t - 1})$ for the updating of the particle filter

ω_{t}^{[n]} = η_{t} \cdot ω_{t - 1}^{[n]} \cdot \exp (- \frac{d^{2} (X_{t}^{[n]}, C_{t}^{\hat{l}})}{2 σ_{d}^{2}})

(10)

Resampling

Finally, the resampling deals with the problem of particle degradation that occurs after several times of iterations. We reproduce a set of particle sets by removing low-weight particles and replicating high-weight particles. The new set of particles are used for the prediction and updating in the next iteration.

Experiment

Experimental setup

We verify the feasibility of our algorithm in an indoor environment. The experimental environment consists of walls, obstacle cartons, and ordinary pedestrians. The measurements are collected using a Scitos G5 robot. The robot is equipped with an Impinj Speedway Revolution R420 UHF RFID reader. The reader provides a reading range up to 7 m. It takes approximately 0.3 s for each inquiry of the RFID measurement. Two antennas (Lairs Technologies SS8688P with an antenna gain of 9.0 dBi) are connected to the RFID reader. The antennas were installed at a height of 1.2 m on the robot and were parallel to the front of the robot. Moreover, the robot is equipped with a SICK S300 laser range finder (with a maximum reading range of 29 m and scan angle of 270°).

During our experiment, a person walked on a rectangular path of 10 × 5 m with a passive tag (Squiggle, Alien Technology, San Jose, CA, USA). This person walked around the rectangular area four times at a speed of 0.4 m/s. In addition, there are multiple static obstacles as well as moving obstacles (human) in the experimental scene. The sampling rate of the RFID reader is set to be 0.4 s per sample. In total, there are a number of 764 measurements recorded during the walking. Furthermore, the robot is fixed at 1 m from the midpoint of the long side of the rectangular path. As shown in Figure 4, there are multiple obstacles around the experiment. At the same time, other pedestrians moved randomly in the area.

Figure 4.

An overview of the experimental setup.

Evaluation of the positioning accuracy under different approaches

In this series of experiments, we used three different approaches to evaluate the tracking accuracy. For the first approach (i.e. velocity matching), we perform velocity matching between the laser cluster velocity and the RFID phase velocity when the RFID information is available. The matching results are used to update the weights of the particle filter. The same implementation is also used by Fu et al.⁴¹ For the second approach (laser-based tracking), we ignored the RFID information and only used laser clustering results for tracking by assuming that the initial location of the object is known. The third approach (i.e. our approach) combines the previous two approaches. The comparison of these three approaches is shown in Table 3.

Table 3.

Comparison of the tracking accuracy under different approaches.

Approaches	Positioning error (m)
Velocity matching	1.33
Laser-based tracking	2.69
Our approach	0.43

For this series of experiments, we set the parameters as $ε = 0.5, ζ = 2, N = 200, d_{\max} = 1.0, K = 4$ , and $σ_{d} = 0.1$ . As can be seen in Table 3, our approach provides a better positioning accuracy when compared with the velocity matching and laser-based tracking. Positioning with a laser range finder gives the worst result. The reason is that laser-based tracking easily fails to track the true object due to the influence of many obstacles in the environment. Figure 5 shows the trajectories using different approaches, and Figure 6 shows the estimated error of different approaches at different timestamps. As can be seen from these figures, the RFID information is only available at approximately 60% locations in the path; therefore, the velocity matching was only performed at these locations, and thus leading to a high positioning error (i.e. 1.33 m). The laser-based tracking method only uses the ranging measurements for positioning. With laser alone, our system is not able to determine the target to be tracked, since there are multiple objects moving in an environment. The target position cannot be directly determined at the beginning of the particle filtering, and we can only rely on the previous position to predict the location of the object at the current time. This is why laser-based tracking method gives worse positioning accuracy than other approaches. When the RFID information is combined, we can determine the position of the moving object from a number of clusters in the environment and therefore improves the positioning accuracy. Our approach (i.e. a combination of velocity matching and laser-based tracking) gives a positioning accuracy of 0.43 m, which is an improvement of 67.6% and 84.1% when compared with the velocity matching and laser-based tracking. The probability of localizing the object within 0.4 m reaches 59.8% (about 457 measurements as compared to a total number of 764 measurements) based on our approach, and only reaches 38.3% (about 293) according to velocity matching. This shows our approach is able to reduce the large positioning error in the velocity matching–based approach by the additional incorporation of the laser ranging information when the RFID readings are not available. Therefore, our approach achieves the continuous tracking of a dynamic object even when the object is out of the RFID reading range.

Figure 5.

Ground truth and estimated trajectory under different approaches.

Figure 6.

Estimation error at different timestamps under different approaches.

In order to verify the robustness of the entire positioning system, we added two dynamic targets carrying RFID tags based on the original experimental scene in Figure 7. In addition, the experimental parameters are set to be the same as the original experimental parameters, and the positioning results are shown in Figure 8. As can be seen from Figure 8, when multiple dynamic tags are used, the overall positioning effect is consistent with the real path and the average positioning error of the three tags is 0.54 m, which is only 0.1 m worse than the positioning error of a single target (24%). The experimental results show that our approach is able to track the positions of multiple objects with a good positioning accuracy.

Figure 7.

Multiple tags experimental environment setup.

Figure 8.

Ground truth and estimated trajectories of multiple tags.

The effect of different antenna configurations

The number of antennas directly affects the detecting range of the RFID sensor. Therefore, in the second series of experiments, we analyzed the effect of different antenna configurations on the positioning accuracy. We fixed $ϵ = 0.5, ζ = 2, N = 200, d_{\max} = 1.0, K = 4$ , and $σ_{d} = 0.1$ . The results are shown in Table 4. From this table, we can clearly observe that the number of antennas has a high influence on the positioning accuracy for both velocity matching–based approach and our approach. We obtain a poor positioning accuracy using only one antenna, since one antenna only covers a small region of the experimental area. We achieved a better positioning accuracy using two antennas. As can be seen from Table 4, our approach reduced the positioning error to 0.43 m using two antennas, which is an improvement of 32.8% as compared to the right antenna alone (0.64 m).

Table 4.

Positioning accuracy under the impact of different antenna combinations.

Configurations	Velocity matching	Our approach
Left antenna	2.52 m	0.69 m
Right antenna	2.34 m	0.64 m
Both antennas	1.33 m	0.43 m

Tracking accuracy under the impact of $ϵ$ and $ζ$ in DBSCAN

Since the DBSCAN clustering is controlled by two parameters radius $ϵ$ and the minimum number of points $ζ$ , it is worth noting that the moving distance of a normal pedestrian in 0.3 s should be between 0 and 1 m. Thus, we set $ϵ$ to be $0.0 \leq ϵ \leq 1.0$ . Meanwhile, an obstacle should contain at least one laser point, that is, $ζ \geq 1$ . Therefore, we performed several experiments to test the effect of $ϵ$ and $ζ$ on the positioning accuracy. We set parameters as follows: $N = 200, d_{\max} = 1.0, K = 4$ , and $σ_{d} = 0.1$ .

Table 5 shows the average number of clusters under different $ζ$ at $ϵ$ . As can be seen from this table, $ϵ$ and $ζ$ has a significant impact on the number of clusters detected. The number of clusters increases with the decrease of $ζ$ at the same $ϵ$ . In addition, the choice of $ϵ = 0.3$ gives more number of clusters as compared with the rest settings of $ϵ$ . Figure 9 visualizes the clustering results for different $ζ$ at $ϵ = 0.5$ . From this figure, the number of clusters decreases as $ζ$ increases. With $ζ = 10$ , there are few clusters that satisfy $N_{ε} \leq 10$ and most of the laser points are considered as noise.

Table 5.

Average number of clusters under the impact of different $ϵ$ and $ζ$ in DBSCAN.

$ϵ$	$ζ$
	1	2	3	4	10
0.0	2.5	2.0	1.9	1.8	1.0
0.05	2.1	1.9	1.5	1.3	1.0
0.1	16.2	9.5	1.6	1.4	1.1
0.2	20.2	17.5	9.4	6.6	1.1
0.3	20.8	15.7	13.2	10.5	1.2
0.4	18.9	15.2	13.8	12.3	1.3
0.5	18.4	15.2	14.1	10.9	3.2
0.6	17.6	14.2	12.1	10.2	3.5
0.7	16.9	13.8	11.9	9.9	4.3
1.0	15.6	13.7	11.7	9.4	6.3

Figure 9.

Clustering results of different $ζ$ at $ϵ = 0.5$ .

The positioning accuracy under the impact of $ζ$ and $ϵ$ are shown in Table 6. As can be seen from Table 6, a setting of $ϵ = 0.5$ and $ζ = 2$ gives the best positioning result. The value of $ϵ$ has a high impact on the positioning accuracy. A too small or too large $ε$ obviously deteriorates the positioning results. A small $ϵ = 0.3$ will segment an object into different clusters, and thus leads to a poor positioning accuracy. With a large $ϵ = 1.0$ , several objects are merged into one cluster, which will also lead to an increase of the positioning error. As can be also seen from Table 6, with a large $ζ$ , two adjacent clusters with high density may be merged into one cluster, which will lead to an increase of the positioning error. A small $ζ$ slightly decreases the positioning accuracy.

Table 6.

Tracking accuracy under the impact of $ϵ$ and $ζ$ .

$ϵ$	$ζ$
	1	2	3	4	10
0.0	5.18	5.09	4.99	5.11	5.43
0.05	4.34	4.88	4.56	4.86	5.14
0.1	0.66	0.63	4.40	4.62	4.68
0.2	0.59	0.54	4.25	4.16	4.57
0.3	0.54	0.52	0.78	1.01	5.86
0.4	0.52	0.49	0.63	0.71	5.78
0.5	0.46	0.43	0.59	0.65	5.15
0.6	0.49	0.46	0.60	0.64	5.51
0.7	0.50	0.47	0.61	0.67	5.23
1.0	0.57	0.54	0.64	0.68	5.83

The effect of different numbers of particles $N$

Since we use the particle filtering for the tracking, the number of particles $N$ directly affects the positioning accuracy and the running time of the algorithm. The results are shown in Table 7. We used a CPU with corei5-7300HQ (2.5 GHz) and 8GB RAM for the processing. The parameters of the experiments are set as follows: $ε = 0.5, ζ = 2, d_{\max} = 1.0, K = 4$ , and $σ_{d} = 0.1$ .

Table 7.

Effect of particle number $N$ on positioning accuracy and running time.

N	Error (m)	Time (ms)
5	2.87	3.85
50	0.51	4.28
100	0.48	4.81
200	0.43	5.82
500	0.43	8.84
1000	0.42	14.16

As can be seen from Table 7, a small $N$ $(N < 100)$ gives an increase of the positioning error, as a small number of particles is not able to well represent the probability density function. The positioning accuracy gets improved with an increasing number of particles, but performing filtering with a large number of particles also consumes more time. With a large $N \geq 200$ , we almost get the same positioning results. Considering the accuracy and time consumption of the algorithm, we choose $N = 200$ for our experiment.

Impact of the range $d_{\max}$

For the laser cluster tracking, the distance that an object moves at two continuous times should be within a certain range $d_{\max}$ . The setting of the range $d_{\max}$ directly affects the correct selection of the cluster. Therefore, we analyzed the accuracy under the impact of range $d_{\max}$ in Table 8. For this series of experiments, we set $ε = 0.5, ζ = 2, N = 200, K = 4$ , and $σ_{d} = 0.1$ . As can be seen from Table 8, $d_{\max} = 1.0$ gives the best positioning accuracy. With a small $d_{\max} = 0.5$ , the approach fails to track the object due to its fast moving speed, thus resulting in a large positioning error (1.39 m). Applying a large $d_{\max}$ (i.e. $d_{\max} > 2$ ) also increases the positioning error due to the inclusion of a large number of unnecessary clusters.

Table 8.

Influence of the value of range $d_{\max}$ on the positioning accuracy.

$d_{\max} (m)$	0.5	1.0	2.0	3.0
Error (m)	1.39	0.43	0.56	0.59

Impact of $K$ best match and translational coefficient $σ_{d}$

$K$ and $σ_{d}$ are the two important parameters for the filter updating. Table 9 shows the impact of the positioning accuracy under these two parameters. The remaining parameters are set as follows $ε = 0.5,$ $ζ = 2, d_{\max} = 1.0$ , and $N = 200$ .

Table 9.

Positioning accuracy under the impact of different $K$ and $σ_{d}$

$K$ (clusters)	$σ_{d}$
	0.01	0.05	0.10	0.50	1.00
1	3.89	2.00	1.44	2.15	4.78
4	0.52	0.47	0.43	0.51	0.59
7	0.54	0.49	0.47	0.52	0.58
13	0.70	0.66	0.49	0.54	0.63

As can be seen from Table 9, $K$ and $σ_{d}$ have a significant effect on the positioning accuracy. A too large or too small $K$ results in a reduction of the positioning accuracy. $K > 1$ leads to a positioning error above 1 m, as there may be more than one moving obstacles in the environment. $K = 13$ will include many unwanted clusters for the updating and reduces the positioning accuracy as well. The best choice of $K$ is $K = 4$ for all our trials. In addition, the positioning accuracy is reduced with a small or large $σ_{d}$ . Given a setting of $k = 4$ , $σ_{d} = 0.1$ gives a positioning accuracy of 0.43 m, which is an improvement of 17.3% and 27.1% as compared to $σ_{d} s = 0.01$ (0.52 m) and $σ_{d} = 1.0$ (0.59 m), respectively.

Conclusion

This article proposed an approach to fuse RFID and laser information to achieve dynamic object positioning using a particle filter. Due to the limited range of the RFID sensor, we use the laser range information to continuously track the position of a moving object. We estimate the phase velocity of the object when RFID measurements are available. This velocity is matched against the laser cluster–based velocity to localize the moving object. If the object is not in the range of the RFID, we only use laser clustering results to track the object. By combining the previous estimated position, we select the closest cluster to update the particle filter. Our approach leads to a positioning accuracy of 0.43 m, which is an improvement of 67.6% and 84.1% as compared to laser tracking and velocity matching, respectively. However, our system will fail to track the moving object when it is occluded by obstacles for a long time. Second, the sampling frequency used in this article is set to 0.3 s. Our approach will fail if the object moves too fast, for example, the moving distance in two continuous measurements is larger than $λ / 2$ (0.4 m), as a phase ambiguity problem will be introduced and our approach is not able to measure the correct radial velocity that the object travels. In the future, we would like to optimize the parameters in the particle filter. For example, the parameter $K$ can be functionalized based on the similarity measures and the number of clusters in an environment. To improve positioning accuracy, another research direction is to integrate inertial sensors for the prediction of the object’s movement.

Footnotes

Acknowledgements

The authors would like to thank Dr Chin Keong Ho from A*STAR, Singapore, for the discussion of the paper.

Handling Editor: Amiya Nayak

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research is supported partially by the National Natural Science Foundation of China (grant nos.: 61601381, 61701421, and 61471306), partially by the China’s 13th Five-Year Plan in the Development of Nuclear Energy (grant no.: 20161295), partially by the Sichuan Province Science and Technology Innovation Seed Project (grant no.: 2018047), partially by Sichuan Science and Technology Program (grant no.: 2019YFH0161), partially by the Postgraduate Innovation Fund Project by Southwest University of Science and Technology (grant no.: 18ycx117).

ORCID iD

Yulu Fu

References

Kjærgaard

Blunck

Godsk

et al . Indoor positioning using GPS revisited. In: International conference on pervasive computing, Helsinki, 17–20 May 2010, pp.38–56. Berlin: Springer.

Yasir

S-W

Vellambi

. Indoor positioning system using visible light and accelerometer. J Lightwave Technol 2014; 32(19): 3306–3316.

Zhuang

Yang

et al . A pervasive integration platform of low-cost MEMs sensors and wireless signals for indoor localization. IEEE Inter Things J 2017; 5: 4616–4631.

Huixia

. The research and application of UHF RFID indoor positioning. Master Thesis, School of Communication and Information Engineering, University of Electronic Science and Technology of China, 2013.

Xie

Yin

Chen

et al . RFID data management: algorithms, protocols and performance evaluation. Chin J Comput 2013; 36(3): 457–470.

Liu

. RFID technical principle and its application. J Centr Univ National 2006; 4: 358–361.

Montaser

Moselhi

. RFID indoor location identification for construction projects. Automat Constr 2014; 39(4): 167–179.

Saab

Nakad

. A standalone RFID indoor positioning system using passive tags. IEEE Trans Indus Electron 2011; 58(5): 1961–1970.

Zhixue

. 2016Research of location-aware method based on RFID technology. PhD Thesis, Changchun University of Technology, Changchun, China.

10.

Bao

Wang

. Random sampling algorithm in RFID indoor location system. In: International workshop on electronic design, test and applications, Kuala Lumpur, Malaysia, 17–19 January 2006, p.6. New York: IEEE.

11.

Huagang

Huang

Gao

et al . Approximate maximum likelihood algorithm for moving source localization using tdoa and fdoa measurements. Chin J Aeronaut 2012; 25(4): 593–597.

12.

Liu

Lau

et al . LANDMARC: indoor location sensing using active RFID. In: International conference on pervasive computing and communications, Fort Worth, TX, 26 March 2003, pp.701–710. New York: IEEE.

13.

Yang

Guo

Liu

et al . Perceiving the slightest tag motion beyond localization. IEEE Trans Mobile Comput 2015; 14(11): 2363–2375.

14.

Joho

Plagemann

Burgard

. Modeling RFID signal strength and tag detection for localization and mapping. In: International conference on robotics and automation (ICRA 2009), Kobe, Japan, 12–17 May 2009, pp.3160–3165. New York: IEEE.

15.

Vorst

Zell

. Semi-autonomous learning of an RFID sensor model for mobile robot self-localization. In: Bruyninckx

Přeučil

Kulich

(eds) European robotics symposium 2008. (Springer Tracts in Advanced Robotics), vol. 44 Berlin; Heidelberg: Springer, 2008, pp.273–282.

16.

Han

Ding

Qian

et al . CBID: a customer behavior identification system using passive tags. IEEE/ACM Trans Netw 2016; 24: 2885–2898.

17.

Arnitz

. Tag localization in passive UHF RFID. PhD Thesis, Graz University of Technology, Graz, 2011.

18.

Povalac

Sebesta

. Phase of arrival ranging method for UHF RFID tags using instantaneous frequency measurement. In: Proceedings of the 20th international conference on applied electromagnetics and communications, Dubrovnik, 20–23 September 2010, pp.1–4. New York: IEEE.

19.

Martínez-Rey

Santiso

Espinosa

et al . Smart laser scanner for event-based state estimation applied to indoor positioning. In: International conference on indoor positioning and indoor navigation, Alcala de Henares, 4–7 October 2016, pp.1–7. New York: IEEE.

20.

Wang

Peng

Ravankar

. A single lidar-based feature fusion indoor localization algorithm. Sensors 2018; 18: E1294.

21.

Zhang

Zhu

Retscher

et al . Three-dimension indoor positioning algorithms using an integrated RFID/INS system in multi-storey buildings. Berlin; Heidelberg: Springer, 2009.

22.

Chae

Choi

Seo

et al . Robust people tracking using an adaptive sensor fusion between a laser scanner and video camera. Int J Distrib Sensor Netw 2013; 9: 521383.

23.

Chen

Liu

Jaakkola

et al . Knowledge-based indoor positioning based on LiDAR aided multiple sensors system for UGVs. In: Position, location and navigation symposium (PLANS 2014), Monterey, CA, 5–8 May 2014, pp.109–114. New York: IEEE.

24.

Zoeke

Ziroff

. Self-localization of distributed radar nodes based on velocity matching of moving targets. In: IEEE radar conference, Cincinnati, OH, 19–23 May 2014, pp.647–650. New York: IEEE.

25.

Ran

Cheng

Wang

. SINS alignment using velocity matching and simplified cubature Kalman filter. Appl Mech Mater 2014; 615: 255–258.

26.

Ester

Kriegel

Sander

et al . A density-based algorithm for discovering clusters in large spatial databases with noise. In: Proceedings of the 2nd international conference on knowledge discovery and data mining (KDD 96), Portland, OR, 2–4 August 1996, pp.226–231. New York: IEEE.

27.

Jiao

Wang

You

et al . Geo-positioning accuracy improvement of multi-mode GF-3 satellite SAR imagery based on error sources analysis. Sensors 2018; 18(7): E2333.

28.

Zhao

Liu

. VIRE: active RFID-based localization using virtual reference elimination. In: International conference on parallel processing, Xi’an, China, 10–14 September 2007, pp.56–63. New York: IEEE.

29.

Hightower

Borriello

Want

. SpotON: an indoor 3D location sensing technology based on RF signal strength. Technical report, University of Washington, Seattle, WA, 2000.

30.

Zhou

Zhang

. Two-dimension localization of passive RFID tags using AOA estimation. In: IEEE international instrumentation and measurement technology conference, Binjiang, China, 10–12 May 2011, pp.1–5. New York: IEEE.

31.

Sarkka

Viikari

Huusko

et al . Phase-based UHF RFID tracking with nonlinear Kalman filtering and smoothing. IEEE Sensor J 2011; 12(5): 904–910.

32.

Kim

et al . TDOA/AOA localization in RFID system using dual indirect Kalman filter. In: IEEE/SICE international symposium on system integration (SII), Kyoto, Japan, 20–22 December 2011, pp.440–445. New York: IEEE.

33.

Liu

Yang

Lin

et al . Anchor-free backscatter positioning for RFID tags with high accuracy. In: IEEE conference on computer communications, Toronto, ON, Canada, 27 April–2 May 2014, pp.379–387. New York: IEEE.

34.

Zou

Zhang

et al . Static map reconstruction and dynamic object tracking for a camera and laser scanner system. IET Comput Vision 2018; 12(4): 384–392.

35.

Lopes

Couto

Jurio

et al . Hierarchical fuzzy logic based approach for object tracking. Knowl-Based Syst 2013; 54(4): 255–268.

36.

Zhou

Ding

. An adaptive online co-search method with distributed samples for dynamic target tracking. IEEE Trans Contr Syst Technol 2017; 26: 439–451.

37.

Xiao

Stolkin

Leonardis

. Dynamic multi-level appearance models and adaptive clustered decision trees for single target tracking. Pattern Recog 2017; 69: 169–183.

38.

Lin

Tsai

et al . Automatic mapping for an indoor mobile robot assistant using RFID and laser scanner. In: SICE annual conference, Takamatsu, Japan, 17–20 September 2008, pp.2102–2108. New York: IEEE.

39.

Gong

Liu

et al . ROS-based object localization using RFID and laser scan. In: International conference on information and automation, Shenyang, China, 6–8 June 2012, pp.406–411. New York: IEEE.

40.

Liu

Xiao

et al . Dynamic object localization based on RFID phase difference and laser ranging information. Chin J Scient Instr 2018; 39(2): 81–88.

41.

Wang

Liu

et al . Moving object localization based on UHF RFID phase and laser clustering. Sensors 2018; 18(3): E825.

42.

Zampella

Jimenez

Seco

. Robust indoor positioning fusing PDR and RF technologies: the RFID and UWB case. In: International conference on indoor positioning and indoor navigation, Montbéliard, 28–31 October 2014, pp.1–10. New York: IEEE.

43.

Retscher

. Using RFID and INS for indoor positioning. Berlin; Heidelberg: Springer, 2009, pp.421–438.

44.

Retscher

. Continuous indoor navigation with RFID and INS. In: IEEE/ION position, location and navigation symposium, Indian Wells, CA, 4–6 May 2010, pp.102–112. New York: IEEE.

45.

Choi

B-S

Lee

J-J

. Mobile robot localization scheme based on RFID and sonar fusion system. In: International symposium on industrial electronics, Seoul, South Korea, 5–8 July, 2009, pp.1035–1040. New York: IEEE.

46.

Xie

Sun

Cai

et al . Tell me what I see: recognize RFID tagged objects in augmented reality systems. In: International joint conference on pervasive and ubiquitous computing, Heidelberg, 12–16 September 2016, pp.916–927. New York: ACM Press.

47.

Liu

Huskić

Zell

. On tracking dynamic objects with long range passive UHF RFID using a mobile robot. Int J Distrib Sensor Netw 2015; 11: 781380.

48.

Duan

Rao

Yang

et al . Fusing RFID and computer vision for fine-grained object tracking. In: INFOCOM—IEEE conference on computer communications, Atlanta, GA, 1–4 May 2017, pp.1–9. New York: IEEE.

49.

Luo

Zheng

et al . An improved dbscan algorithm to detect stops in individual trajectories. ISPRS Int J Geo-Inf 2017; 6(3): 63.

50.

Zhang

Marsic

et al . Online people tracking and identification with RFID and Kinect, 2017, https://arxiv.org/abs/1702.03824

Continuously tracking of moving object by a combination of ultra-high frequency radio-frequency identification and laser range finder

Abstract

Keywords

Introduction

Related work

RFID-based approach

Dynamic target-tracking approach

Fusion with laser ranging information

Fusion with other sensors

Overview

Laser clustering based on DBSCAN

DBSCAN clustering of laser ranging data

Estimate moving velocity of a cluster

Velocity matching

RFID phase velocity estimation

RFID and laser velocity matching

Tracking the dynamic object based on the particle filter

Prediction

Update

Resampling

Experiment

Experimental setup

Evaluation of the positioning accuracy under different approaches

The effect of different antenna configurations

Tracking accuracy under the impact of ϵ and ζ in DBSCAN

The effect of different numbers of particles N

Impact of the range d max

Impact of K best match and translational coefficient σ d

Conclusion

Footnotes

Acknowledgements

Declaration of conflicting interests

Funding

ORCID iD

References

Tracking accuracy under the impact of $ϵ$ and $ζ$ in DBSCAN

The effect of different numbers of particles $N$

Impact of the range $d_{\max}$

Impact of $K$ best match and translational coefficient $σ_{d}$