Night vehicle target recognition based on fuzzy particle filter

Abstract

All along, the identification of night-driving safety car features is a major research direction in the field of intelligent traffic management, with a wide range of applications and development space, and these identification technologies include theoretical knowledge and important theoretical research in many fields. Due to the interference of lights and other light sources, the gray value of the background area also changes frequently. A common method during the day is to detect these background areas as moving vehicles, which greatly reduces the detection accuracy. The most reliable information at night is the headlights. If the light can be accurately detected and other sources are excluded, accurate detection can be performed and tracking accuracy can be guaranteed. Vehicle safety assisted driving technology is one of the main research directions of intelligent transportation systems. It mainly uses computer technology, sensor technology and communication technology to collect and analyze the state information of roads, other vehicles and drivers. Provide advice and warnings to the driver before reaching the danger, determine current traffic conditions and avoid traffic accidents in advance. This paper studies some problems of night vehicle target recognition and detection, mainly the division of target and background, and the classification and recognition of target extraction. To solve these problems, a particle filter algorithm is introduced to introduce nonlinear statistics. The fuzzy theory is introduced to classify the video processed by the particle filter algorithm. The target recognition is realized by the method, and the purpose of identifying the night vehicle target is achieved. Comparative experimental analysis shows that this method is more accurate and powerful than the common target recognition algorithm and can be applied to real scenes.

Keywords

Night vehicle recognition particle filter algorithm nonlinear statistics fuzzy clustering

1 Introduction

Road traffic safety has always been a worldwide problem. With the increase in the number of private cars in recent years, the safety of vehicle driving has become more and more important. Since vision is the most important way for human beings to perceive the world, and visual information describes people’s information, scientists are trying to develop a new science, computer vision [1 –3], by providing human visual ability to machines to improve machine intelligence. Visual-based Vehicle Safety Driving Assistance System [4 –6] uses a camera to capture traffic scenes around the vehicle, then analyzes image sequences, extracts relevant information, and finally executes appropriate judgment and decision-making control. Therefore, the system can correct the driver’s decision-making errors and incorrect operation control to a certain extent, and reduce the occurrence of traffic accidents.

Visual-based Vehicle Safety Driving Assistance System uses sensors to predict road hazards in advance from an active safety perspective. It reminds you of traffic accidents when you approach hazards and do not take necessary measures to protect vehicles. Safety of drivers and pedestrians ensures smooth roads and reduces economic losses. Therefore, the research of vision-based vehicle safety driving assistant system is of great significance to traffic safety, road smoothness and social and economic development.

Particle swarm optimization (PSO) [7 –9] was proposed by Eberhart, an American electrician, and Kennedy, a social psychologist in 1995. It is a method of collective wisdom, inspired by social behavior in life. It is on this basis that when natural organisms take social actions, they will form the exchange and sharing of information within groups. These collective actions are very beneficial to the evolution of the whole population. Collaboration among groups helps to find the best region in the global space on which the PSO algorithm is based.

In this paper, we propose a target recognition algorithm based on fuzzy particle filter, which uses computer simulation technology to realize. In the first part of this paper, the research topic is traffic image data collected through fixed installation of the camera. Through analyzing and processing the image, the vehicle can be extracted and recognized in real time. Other applications are based on different experimental data. The experimental data and research analysis show that the method of combining brightness difference and extracting the connecting area of video image can be detected and found in the night scene to detect and locate the headlights of mobile vehicles. Fixed and road-driven vehicles can effectively eliminate the influence of light from taillights, and calculate traffic flow in images based on the results of detection and location. The rest of the paper is organized as follows: Section II introduces the related work of the article. Section III introduces the fuzzy clustering algorithm [10 –12], particle swarm optimization algorithm and the fuzzy particle filter algorithm. Section four introduces the introduction of the proposed fuzzy clustering and particle swarm optimization algorithm to the night vehicle recognition algorithm, and obtains the results of the introduction and summation of the algorithm. The Section V summarizes all the work of the paper. In this paper, a new algorithm based on fuzzy particle filter is proposed. This algorithm can achieve a very good effect on night target recognition. Its recognition effect can reach 93%, and it has a high recognition rate. It can be applied to real life.

2 Related work

The most commonly used methods for target recognition are Bayesian filtering based on two methods [13 –15]: Kalman filtering method [16 –18] and particle filtering method [19 –21]. Particle filter algorithm uses the idea of Monte Carlo method to approximate the posteriori probability of the system using samples or particles, and applies them to the state estimation of non-linear systems. Particle filter algorithms have been widely used in the field of video object tracking, because they can handle all kinds of probabilities and are not limited to linear Gauss distribution probabilities. Different types of particle filters are also proposed and applied, such as particle filter algorithm based on kernel density estimation, particle filter algorithm based on Gabor feature, and particle filter algorithm based on Haar feature.

Since the 1980 s, people began to apply image and video processing technology to dynamic vehicle detection and tracking algorithm. Recently, domestic and foreign scholars have made extensive research on video-based vehicle detection technology, and put forward various algorithms in-depth study. However, most algorithms are implemented under normal daytime lighting conditions, and some algorithms are processed under special weather conditions, such as fog, rain, snow, etc. The study of night conditions is relatively few, and the effect is not very ideal. With the rapid development of surveillance equipment, the shooting effect of traffic video is more and more obvious, and the processing ability of computer is constantly improving. Vehicle detection technology based on video has become a research hotspot in the field of intelligent transportation system. Compared with other detection technologies, vehicle detection technology based on night video has the advantages of convenient installation and maintenance, no traffic interruption, low cost, large amount of information analysis, no impact on road life, integrated manual and automatic monitoring, etc. At present, the development speed is very fast.

3 Fuzzy clustering and particle filtering

3.1 Fuzzy cluster analysis

Clustering analysis is an unsupervised data analysis technology, which does not require training in classification algorithms. Clustering is to divide data object groups into different classes or clusters according to some criteria to achieve the goal. That is to say, the similarity between data objects of the same class is as large or as small as possible. The similarity between data objects that are not in the same class is as small or as large as possible. The principle of clustering is to collect as many data objects of the same class as possible and keep different data objects as much as possible. With the increasing application of fuzzy clustering, relevant experts and scholars have also strengthened the research of fuzzy clustering algorithm. Among many fuzzy clustering algorithms, the fuzzy C-means clustering algorithm is the most widely used. Fuzzy C-means clustering algorithm calculates the membership degree of each sample point of all class centers by objective function, and determines the universality of the sample points according to the membership degree calculated, so as to achieve the purpose of automatic classification of sample data.

Fuzzy C-means clustering algorithm divides the set of data samples containing N data into C-class and X-expression as (1) $X = {x_{1}, x_{2}, \dots x_{n}}$ (1)

The general definition of objective function is

$J (U, C) = \sum_{j = 1}^{c} \sum_{i = 1}^{n} μ_{ij}^{α} d_{ij}^{2}, 1 ⩽ α ⩽ \infty$ (2)

The U = [μ_ij] _c×n representation is the membership matrix. The degree of membership represents the degree to which the first sample point belongs to the j th clustering center, and in the clustering algorithm based on objective function, the clustering center between particles is usually determined by calculating the distance between the particles and the clustering center. The smaller the contribution is, the closer the particles are to the center of the cluster, and the greater the degree of this category. According to the normalization rules, the membership of data samples must satisfy the sum of 1 that is the formula. $\sum_{j = 1}^{c} μ_{ij} = 1, \forall i = 1, 2, 3, \dots n$ (3)

In formula (2), A is a weighted exponent, which is usually taken as 2, D is the Euclidean distance between the first sample point and the j cluster center, and its expression is as follows. $d_{ij} = ∥ c_{j} - X_{i} ∥$ (4)

The optimization process is a simple iteration process. It starts from a random initial clustering center. By calculating the minimum value of J_c, it adjusts the clustering center and the membership degree μ_ij of each sample point, converges at the local minimum point of J_c, and determines the classification. Assuming that a data set contains n data objects, the principle of partitioning method is to construct C partitions (c<), a data set contains n data objects, and the principle of classification method is to construct the (c<) classification method is to set the number of sets in advance and to classify them initially. By updating the database to move the data in the same group to make the data target similar, there are great differences between different groups. Mountain climbing gives search precedence over the extreme majority of local methods using heuristic techniques and finding complex centers. The segmentation method is suitable for older data aggregation, but it can produce different set results with different initial values. There are two clusters: a large application cluster based on Randomized Search and Clustering Application. The block diagram of partition clustering algorithm is shown in Fig. 1.

Fig. 1

Advantages and disadvantages of binary encoding and floating-point encoding.

The hierarchical clustering method divides the data set into groups (classes) to form a clustering tree. According to the clustering method, the data set can be divided into top-down hierarchical clustering and bottom-up hierarchical clustering. Hierarchical clustering is summarized. Compressive hierarchical clustering means that each data object is first processed into a class, then merged step by step until a Non-Merging set is formed. Hierarchical disassembly clustering treats the whole data object as a class, and then creates several subclasses until a given rule is partitioned and clustered step by step. Typical hierarchical clustering algorithms are BIRCH (Balanced Iterative Reducing and Clustering using Hierarchies), CURE (Clustering Using Representatives) and CHAMALEON. Hierarchical clustering method does not need the preset cluster number of the cluster, nor does it need the initial partition, so it does not need the current state of the cluster data set. The bottom-up hierarchical clustering method combined with cyclic relocation is usually used to achieve better clustering results. Figure 2 is a schematic diagram of clustering of data sets a, b, c, d, e, showing the clustering and splitting process.

Fig. 2

The main computational steps of the genetic algorithm.

3.2 Particle swarm optimization

PSO arithmetic comes from the study of birds’ predation behavior. The simplest and most effective way to find food when birds eat is to find cities around the birds closest to food. Inspired by the behavioral characteristics of this biological population, PSO algorithm is used to solve optimization problems, where each particle in the algorithm represents the potential solution of the problem, and each particle corresponds to the fitness value determined by the number of physical surfaces. Particle velocity determines the direction and distance of particle motion, and adjusts the velocity dynamically according to the movement experience of itself and other particles, so as to realize individual optimization in the distinguishable space. The PSO algorithm first initializes the particle group in the feasible solution space. Each particle represents the optimal solution of the extremum optimization problem. Particle attributes are represented by three indexes: position, velocity and fitness. The fitness function is calculated. The value of the fitness function represents the advantages and disadvantages of the particle. Particle motion in the solution space tracks individual extreme Pbest and extreme population Gbest to update individual positions. Extreme Pbest is the best location of fitness values calculated from the location of individual experience, and extreme Gbest is the best location of fitness retrieved by all particles in the population. The fitness values of the new particles are calculated each time the particles are updated, and the positions of individual extreme value Pbest and global extreme value Gbest are updated by comparing the fitness values of the new particles with the fitness values of individual extreme value and group extreme value.

PSO (Standard Particle Swarm Optimization) algorithm is a swarm intelligence optimization method based on bird feeding development. As shown in Fig. 3, only A found food traces in birds consisting of A to E, so the position of A bird is the best position in the chicken flock. Another bird also needs to enter the bird’s position, and the bird needs to get food. This is the most primitive logic in particle swarm optimization. The principle of standard particle swarm optimization is shown in Fig. 3.

Fig. 3

Visual imaging technology.

For a single particle, it uses its own information and other particle information to update its location in order to achieve the goal of approaching food and reaching the optimal location. Before updating the location, it is necessary to first evaluate the existing particles, select their own optimal Pbest after several generations of updating, and then select the global optimal individual Gbest after updating the overall generation. Then, for a single particle, there are three updating parameters, including the final updating rate, the best individual lead EST and the best global Gbest. Particle position updating will generate new particles by referring to these three vectors and appropriate weighting factors to complete particle position updating.

3.3 Particle filter algorithm

Particle filter is a state estimation algorithm for non-linear non-Gaussian systems. It is a recursive Bayesian estimation based on Monte Carlo simulation. Particle filter is not constrained by system linearity and Gaussian noise, which determines its wide range of applications. It has certain advantages in target tracking, image processing, data prediction and signal detection. Next, we introduce Bayesian estimation, Monte Carlo sampling and sequential importance sampling. Finally, we introduce resampling technology to explain the basic principle of particle filter step by step.

The state space model of the system is a time domain model describing the relationship between the system state and the observation, and is widely used in the field of optimal estimation. The state space model of the system is described by the state space representation, which consists of the state equation representing the state of the system as a time function, and the observation equation representing the relationship between the state and the state observation of the system. $x_{k} = f (x_{k - 1} + w_{k})$ (5) $y_{k} = h (x_{k}) + v_{k}$ (6)

Among them, x_k ∈ R^{n
_x} is the state variable at time, y_k ∈ R^{n
_y} is the observed value at k time, f(·) and h(·) are the state transition function and measurement function respectively, w_k and v_k are the system noise in the same dimension as x_k and the measurement noise in the same dimension as y_k. In particle filter, the above state space model is described in a statistical way as p (x_k|x_k - 1) (state transition probability density, obtained from state equation) and p (y_k|x_k) (state observation likelihood probability density function, obtained from measurement equation). At the same time, it is assumed that the system state obeys the first-order Markov process, that is, the state x_k of k-time is only related to the state x_k - 1 of its previous time. The state space model is shown in the following Fig. 4:

Fig. 4

State-space model.

Monte Carlo method presupposes sampling from probability density function, but it is difficult to extract samples directly from posterior distribution p (x_O:k|y_1:k). Therefore, in particle filter, an important probability density function (or recommended distribution function) q (x_O:k|y_1:k), which is easy to sample, is introduced.

$E (f (x_{O : k})) \approx \sum_{i = 1}^{N} f (x_{O : k}^{i}) w_{k} (x_{O : k}^{i})$ (7)

Every time a new observation value y_k is introduced, new samples are extracted from the proposed distribution q (x_O:k|y_1:k), and the importance of each sample is recalculated. As time goes on, more and more storage space is occupied. Therefore, a sequential importance sampling (SIS) method is proposed, which realizes importance sampling recursively, updates weight recursively and estimates importance probability density function. First, the distribution of recommendations can be divided into: $q (x_{O : k} {| y}_{1 : k}) \approx q (x_{k} {| x}_{O : k - 1}, y_{1 : k}) q (x_{O : k - 1} {| y}_{1 : k - 1})$ (8)

As time goes on, the variance of importance weight in SIS algorithm will increase. After repeating a group of particles several times, there are only a few or even larger particles. Because it tends to be small and weighted slowly to zero, the number of particles that can be used to estimate the probability of state decreases sharply, which leads to the particle decomposition problem in SIS algorithm. In order to alleviate the particle decomposition in SIS algorithm, the resampling process is introduced. After repeated sampling, the discrete density N generates a new group of particles instead of the original group of particles, in which N becomes the number of particles and new particles. It has the same weight. The resampling process reduces the variance of weights by avoiding a large amount of computation of small weighted particles that have little effect on the contribution of the backward probability density function.

Selecting the critical probability density function as the prior probability density and introducing the resampling step in the standard SIS algorithm as mentioned above is also called sequential importance resampling sequence importance sampling. The particle decomposition problem in this method, i.e. the existence of random noise and the passage of time, results in the large dispersion of particle weight, more and more particles deviate from the reverse distribution of the system state, and the weight of these particles changes. It is a very small and negligible level, so the large weight is concentrated on a very small number of particles, so the post-probability wastes a lot of time on the particles and the effect is very small, so it can not accurately describe the post-probability density function of the state. Choosing proper importance probability density function is a method to solve this problem.

4 Experimental design and analysis

4.1 Experimental example

Firstly, the night traffic video image is preprocessed. The preprocessing involves the segmentation of the region of interest and the bi-value segmentation of the vehicle ramp, and the relevant rules and strategies are formulated based on the statistical analysis of the geometric and optical information of the scene. Develop and use this rule to determine all candidate lights that will provide you with light test results. Fast and effective optical detection is very important for follow-up tracking. Compared with the training classification method, the above method requires much less sample size than the training classifier, and the process is simple and expensive less time. Solve the problems of matching light pairs, night vehicle tracking, short-term vehicle occlusion and complex vehicle type resolution. Tracking and matching headlights are divided into two main processes: lighting tracking and matching. In the tracking method, illumination tracking and matching provide mutual support information. The motion information obtained through lamp tracking will provide a reliable basis for lamp matching, and in the process of tracking, the corresponding matching lamp situation information is conducive to improving the accuracy of tracking. In addition, we propose a bi-directional predictive trajectory matching technique based on velocity estimation to continue to predict vehicle trajectories. This can solve short-term obstacles and inaccurate tracking problems caused by noise detection.

Initial phase: This phase requires manual tracking of the target and allows the program to calculate the tracking function. For example, the color features of the target can be used. Specifically, starting with code, the object feature of the artificial mouse is that the program automatically calculates the rectangularity of tones in the area, because the rectangle can be expressed as a vector, so the object feature is a vector of N* 1. Fifth, the search phase: What are you looking for particle consequences. First, you need to inject particles. Two of the most commonly used methods. That is, 1) uniformly (particles scattered uniformly across the image plane), 2) can be understood by placing Gothic distribution near the target of the frame and putting it closer to the target so that it is farther away from the eye level. The use is a backward method, and the target features (hue rectangle, vector V) obtained in the initialization step use the position of particles to retrieve the target, in order to use the color features of the picture to obtain hue straightness, this orthogonality calculation is similar to the orthogonality of the target. Similarity varies. The simplest way is to calculate the sum: each particle has a synchronization rate, and then add the synchronization rate again and again. According to the synchronization ratio, we make the marginal mean. The synchronization rate of this report is Wn. The target is the most possible pixel coordinates x = sum (x_n * w_n), y = sum (y_n * w_n). Let go and place fewer particles at a lower synchronization rate. 2 > 3>4 > 2, repeat the movement of the target repeatedly.

4.2 Night vehicle recognition

Based on the system equation and observation equation, the vehicle position prediction is updated in the next frame. Perform target parameter matching between consecutive frames. First, the characteristic parameters of the headlamp in the current frame are stored in the set S = {B_i, i = 1, 2, 3 ⋯ N}, and then the known headlamp B_i is the headlamp in the current frame. The most similar target is determined according to the similarity, the known target set S is updated, and the processing is performed frame by frame according to the video sequence to complete the parameter matching. Kalman filter is used to effectively predict the position of the known target in the current frame and track the position of the target at the next moment to track the temporary loss of the target. Combined with parameter matching and Kalman filter, the target tracking stage is as follows: (1) from the first frame of video, the target in the scene is extracted to its initial value, and the fuzzy particle filter is used to predict the target position in the next frame. (2) Compare the predicted value with the target extracted from the current frame. If the match is made, go to step (4), and if it fails, go to step (3). (3) Check whether the exception is the occurrence of a new target, the disappearance of the target or the termination of the target. (4) Update the model and go to step (2) to start the loop. According to the detection and continuous tracking of the target, the moving speed and position of the target can be obtained. If the position of the target in the next frame can be predicted, the target can be located, then matching features and predicting the tracking yes. You can not only reduce the tracking speed to narrow the search range, but also eliminate unpredictable locations, so as to improve tracking accuracy and make tracking more accurate.

Because the brightness of the lamp is close to 255 and the background brightness is close to zero, the part below the brightness TL is not considered when calculating the threshold value. In order to obtain the brightness value TL, the headlamp images collected from 200 different night high-speed scenes are shown in the statistical brightness histogram as shown in Fig. 6. The horizontal coordinate is the brightness value and the vertical coordinate is the number of pixels. It conforms to the Gauss distribution x to N ∼ (μ, σ²), the mathematical expectation μ and the standard deviation σ² in the histogram, fits the Gauss function, and takes μ - 3σ as the brightness of the lamp. The value TL, TL = 168, was found by Gauss fitting. Statistical headlamp brightness is shown in Fig. 7. The matching of headlights can be divided into three stages. The first stage is to use headlights in the same horizontal direction as the set S_y. Another group of ${BP}_{i}^{k}$ (k is the number of frames, i is the index of headlamp group), and unsuccessfully matched lights are used as groups. The third step is to match multiple headlights ${BP}_{i}^{k}$ belonging to the same vehicle with this group. In this paper, the matching of lights must satisfy the limitations of distance and size. For several lights of the same car, the coordinates of B_i are similar and the area difference is very small.

Fig. 5

Target tracking results.

Fig. 6

Statistical histogram of headlight brightness.

Fig. 7

Fitness and fitness values.

4.3 Experimental results

For videos captured in the above environment, even if the video has a large amount of traffic, low light and multi-light, the method in this white paper also has a good detection effect, with a very low false alarm rate and a low detection rate. It also runs at a speed of about 65 FPS (frames per second) to meet real-time requirements. At the same time, the proposed fuzzy particle filter algorithm is used to detect the above video. The experimental results are shown in Fig. 8.

Fig. 8

Detection accuracy.

The results show that the Opportunity Presentation Fuzzy Particle Filtering algorithm has lower false alarm rate and missed detection rate. In traffic jams, the road reflection time is longer and often considered as the headlight, so the false detection rate is slightly higher than that of the other two guns, and because multiple vehicles light up this situation, the evidence of the false detection rate in the lane is very high. In this specification, the manual selection 250 in this specification cannot identify the type of vehicle with the type of sample image and the type of vehicle with the positive and negative training sample image, and each test image (50) can be identified. SVM classification, classification type and non-recognition of feature vectors extracted from all samples. At present, there are four linear kernels in SVM classifier kernel, which use RBF radial basis function, polynomial function and sigmoid function. Because the selected kernel function is not directly based on, each type of kernel function is good, and paper linear kernel is chosen. The larger the output value of SVM, the higher the image complexity. As shown in Table 1, the first set of the third set of data shows that the correlation between energy sets monotonously increases and decreases the edge monotonous entropy contrast.

Table 1

Complexity statistics of experimental images

Picture	Power	Entropy	Contrast ratio	Relevance	Edge contrast
a	0.15	3.2	1.74	0.1	0.08
b	0.11	3	0.94	0.08	0.07
c	0.12	2.7	0.77	0.09	0.07
d	0.09	3.3	1.85	0.08	0.09

The test set in Fig. 9 has two groups. The first group is a set of vehicle images that can be visually recognized. The classifier can classify 96% of the samples into recognizable groups, while the second group can not be seen. For image sets, the classifier can classify 94% of the samples into unrecognizable sets. The accuracy of the two experiments is so high that recognizable and unrecognizable vehicle images can be distinguished effectively. Therefore, four features and edge ratios obtained from gray level co-occurrence matrix are used to explain the complexity of vehicles. This is possible. Analyse whether the vehicle type is identifiable. For recognizable vehicles, BOF algorithm is used for recognition, and for unrecognizable vehicles, the physical characteristics of lights are used for judgment.

Fig. 9

Classification results of image complexity.

5 Conclusion

The detection of moving objects in video images has become a well-known direction in computer vision. It has broad application prospects in intelligent traffic management, safety monitoring and automation. Researchers began to study their skills. The motion detection technology based on video image integrates the related technologies in the field of digital video image processing and artificial recognition. Video image is the subject of analysis. By analyzing the image of the set area, we can get the motion information, and use digital image processing technology to remove many natural and artificial. In order to obtain interference and accurate and effective detection results, the basic content of moving object detection based on video image is to extract moving objects from continuous video sequence images, and identify and track the extracted moving objects. Compared with the traditional vehicle detection technology, it has the following advantages: simple installation and maintenance, only need to install the camera on the road, can obtain detailed traffic information data, such as traffic information, speed, vehicle type and so on.;

Based on the investigation of the related work at home and abroad, the vehicle detection in night traffic video is carried out based on the characteristics of vehicle lights. A method of vehicle target recognition at night based on fuzzy particle filter is proposed. The basic theory of the method of tracking dim and small targets before detection is introduced in detail. The specific realization process of vehicle target detection and the different problems solved by the two methods in their respective fields are described. The advantages and disadvantages of the two methods in target detection and tracking are compared simply. Secondly, the basic theory of the fuzzy particle filter algorithm is introduced in detail. Then on this basis, the vehicle lights are tracked and matched, and the robust night vehicle tracking is experimented. The accuracy of the method is verified through the verification of the detection, accuracy and reliability of multiple vehicle images.

Footnotes

Acknowledgments

This work was supported by National Natural Science Foundation of China under Grant 61703280.

References

Voulodimos

, Doulamis

and Doulamis

, Deep Learning for Computer Vision: A Brief Review, Computational Intelligence and Neuroscience (2018), 1–13.

Barré

, Stöver

B.C.

and Kai

F.M.

, LeafNet: A computer vision system for automatic plant species identification, Ecological Informatics 40 (2017), 50–56.

Ioannidou

, Chatzilari

and Nikolopoulos

, Deep Learning Advances in Computer Vision with 3D Data: A Survey, Acm Computing Surveys 50(2) (2017), 20.

, Choi

and Park

, In-depth understanding of lane changing interactions for in-vehicle driving assistance systems, International Journal of Automotive Technology 18(2) (2017), 357–363.

Böhmländer

, Dirndorfer

and Albayatti

A.H.

, Context-aware system for pre-triggering irreversible vehicle safety actuators, Accident; Analysis and Prevention 103 (2017), 72.

Yue

, Abdel-Aty

and Wu

, Assessment of the safety benefits of vehicles’ advanced driver assistance, connectivity and low level automation systems, Accid Anal Prev 117 (2017), 55–64.

Chatterjee

, Sarkar

and Hore

, Particle swarm optimization trained neural network for structural failure prediction of multistoried RC buildings, Neural Computing & Applications 28(8) (2017), 2005–2016.

Wang

, Jin

and Doherty

, Committee-Based Active Learning for Surrogate-Assisted Particle Swarm Optimization of Expensive Problems, IEEE Trans Cybern 47(9) (2017), 2664–2677.

Fei

L.I.

, Liu

J.C.

and Shi

H.T.

, Multi-objective particle swarm optimization algorithm based on decomposition and differential evolution, Control & Decision 32(3) (2017), 403–410.

10.

, Ma

and Wang

, An Improved Particle Filter Algorithm Based on Ensemble Kalman Filter and Markov Chain Monte Carlo Method, IEEE Journal of Selected Topics in Applied Earth Observations & Remote Sensing 8(2) (2017), 447–459.

11.

Cui

, A New Method for State of Charge Estimation of Lithium-Ion Batteries Using Square Root Cubature Kalman Filter, Energies 11(1) (2017), 209.

12.

Cho

, Kim

E.K.

and Jang

, Improved positioning method for Magnetic Encoder type AGV using Extended Kalman Filter and Encoder Compensation Method, International Journal of Control Automation & Systems 15(4) (2017), 1844–1856.

13.

Kutschireiter

, Surace

S.C.

and Sprekeler

, Nonlinear Bayesian filtering and learning: A neuronal dynamics for perception, Sci Rep 7(1) (2017), 8722.

14.

Berg

, Ende

S.V.

and Lammering

, Calibration of potential drop measuring and damage extent prediction by Bayesian filtering and smoothing, International Journal of Fatigue 100 (2017), 337–346.

15.

El-Shafai

, El-Rabaie

and El-Halawany

M.M.

, Recursive Bayesian Filtering-Based Error Concealment Scheme for 3D Video Communication Over Severely Lossy Wireless Channels, Circuits Systems & Signal Processing 37(20) (2018), 1–32.

16.

, Ma

and Wang

17.

Cui

, A New Method for State of Charge Estimation of Lithium-Ion Batteries Using Square Root Cubature Kalman Filter, Energies 11(1) (2018), 209.

18.

Cho

, Kim

E.K.

and Jang

19.

, Ma

and Wang

20.

Aykaç

, Erçelebi

and Baha Aldin

, ZigBee-based indoor localization system with the personal dynamic positioning method and modified particle filter estimation, Analog Integrated Circuits & Signal Processing 92(3, 4) (2017), 1–17.

21.

Xiang

, Bayesian Signal Processing: Classical, Modern, and Particle Filtering Methods, 2nd Ed, Journal of the Acoustical Society of America 141(2) (2017), 779.