Overcoming Challenges of Distributed Fiber-Optic Sensing for Highway Traffic Monitoring

Abstract

This work presents a wide-area highway monitoring system based on distributed fiber-optic sensing (DFOS) as a cost-effective way of gathering traffic information at numerous sensing points along a fiber cable. The primary advantage of our proposed DFOS system is that it utilizes an existing fiber cable buried beneath the highway to detect and localize the vibrations of passing vehicles. Each section along the fiber cable acts as a sensing point and registers the vibration of nearby vehicles. The amplitude and location of vibrations as measured by DFOS can supplement the information obtained from existing point sensor systems (traffic camera, inductive loop detector) which are typically installed hundreds of meters apart. We trained a neural network for speed estimation (SpeedNet) and also proposed novel solutions to some of the challenges posed when using DFOS to monitor traffic. To demonstrate the potential of DFOS, we conducted a field test for two days on a 45-km section of the Tomei expressway in Japan. Our proposed SpeedNet model estimated average traffic speeds every minute for an overall accuracy of over 90% as compared with existing loop detector-based sensors. As cameras suffer from weather and intensity changes, and loop detectors can be difficult to install at multiple locations, monitoring traffic using DFOS over an existing fiber-optic cable shows tremendous potential.

Wide-area traffic monitoring is an important function in every country. The effectiveness of this monitoring plays a vital role in determining the response time to accidents, congestion, and so forth. Traditional monitoring systems are an aggregate of many sub-systems that analyze the data obtained from point sensors such as traffic cameras, inductive loop detectors, LIDAR-guns, and so forth. ( 1 – 4 ). These sensors are strategically placed over an entire target region, thereby enabling the Traffic Management Centers (TMCs) to operate effectively. Although these systems are difficult to both install and maintain, they are essential for the smooth functioning of large metropolitan areas. However, to effectively monitor long highways for congestion and accidents, a sensor has to be installed at least every 500 m ( 5 ). So, when monitoring a highway of 50 km, we need at least 100 sensors, a means of real-time data transmission from each sensor, a sub-system analyzing each sensor’s data, and finally a central system which manages all these sub-systems. Although the advent of technology has eased this complex process, the cost of installing and maintaining it is still demanding.

A cost-effective solution which can be integrated easily into existing monitoring systems is a single distributed fiber-optic sensor based monitoring system as shown in Figure 1. A key advantage of distributed fiber-optic sensing (DFOS) is that every small segment (of a few meters) along an optical cable is a potential vibration sensor ( 6 ). When an optical pulse is transmitted from one end of the cable, scattering occurs at each small segment and generates back propagating light. Vibration amplitude acting on each segment can be measured by analyzing the corresponding back propagated light. Using this principle, one can perform distributed sensing along an existing fiber cable (laid beneath a highway) to measure the vibration exerted by vehicles traveling in proximity to the cable ( 7 ). As DFOS essentially measures the vibrations of passing vehicles, it is possible to detect traffic flow rates, average speeds, and the number of vehicles. These can then be used to detect congestion, abnormal traffic patterns, and travel times. The potential of DFOS for measuring the traffic properties has recently being studied ( 6 – 10 ). To the best of our knowledge, these is no literature which provides sufficient information on how the sensor data is processed and used to monitor traffic.

Figure 1.

A typical distributed fiber-optic sensing (DFOS)-based highway traffic monitoring system.

The primary motivation of our work is to provide a means of processing and modeling the data measured by DFOS for estimating average traffic speeds. We propose a speed estimation model ‘SpeedNet’ which is engineered based on the popular convolutional neural network architecture. We hope that our model can serve as a fundamental framework on which future DFOS based traffic monitoring applications can be designed. In the remainder of this work, we first discuss the operating principle of DFOS and its advantages over traditional sensors, followed by the relevant literature. We then present our proposed methodology for using DFOS for traffic monitoring, the challenges of using DFOS data effectively, and discuss ways of overcoming them. We then show the results of a field test done on the Tomei expressway in Japan where DFOS data was collected for 2 days over 44.5 km. Finally, we trained and tested the SpeedNet model and compared the predicted traffic speeds with those of traffic cameras and loop detectors. The final sections include discussions and conclusions.

Distributed Fiber-Optic Sensing

DFOS is based on optical time-domain reflectometer (OTDR) technology ( 11 ). The system injects light pulses from one end of the cable, and each light pulse coherently interacts with the numerous scattering points along the cable. In the presence of an external vibration at some point on the cable, the fiber’s length and refractive index change slightly, thereby altering the characteristics of the corresponding backscattered light ( 12 ). During scattering, molecules are raised to a virtual energy state and later release light as they fall back to their original energy state. When this light returns to the cable’s start point, the amplitude and location of vibrations are decoded.

Distributed sensors provide numerous advantages in addition to being economical. In contrast to widely popular point sensors like traffic cameras and loop detectors, a single DFOS can cover a wide area and does not need a separate means for sensor data transmission. They are also less sensitive to changes in the weather, unlike cameras ( 1 , 2 , 13 ). As we can use the network of fiber cables already laid beneath most roads and railway structures, we do not need a difficult installation procedure, unlike loop detectors ( 3 , 14 , 15 ). In addition, a traffic monitoring system based on DFOS is extremely power efficient as it requires only a limited power to operate the DFOS in contrast to roughly a hundred traditional sensor sub-systems ( 2 , 16 ). They are also less prone to physical damage and electro-magnetic interference ( 12 ). Although distributed sensing using a fiber-optic cable can also measure changes in the environment like temperature, acoustics, and so on ( 11 , 17 ), we will limit our focus in this paper to sensing vibrations.

There is an enormous amount of literature on the use of traffic cameras and loop detector based systems for traffic monitoring ( 1 , 14 – 20 ). These systems are used extensively by TMCs to monitor abnormal traffic patterns, accidents, and congestion for rerouting traffic and responding to disasters ( 21 , 22 ). Although there is an extensive amount of literature on the applicability of DFOS for monitoring railway traffic ( 23 , 24 ), gas flow in pipelines ( 25 , 26 ), intrusions ( 12 , 27 ), and structural health ( 28 – 30 ), the use of DFOS to monitor traffic is still an emerging field ( 9 ). This motivates us to introduce a basic framework required to realize a DFOS based highway traffic monitoring system.

Highway Traffic Monitoring using DFOS

Monitoring highways poses a significant challenge as they cover long distances and it is often uncertain as to where a traffic accident, congestion, or both, will occur. DFOS can overcome this challenge as it is able to measure the vibrations of passing vehicles along the entire length of a buried fiber cable. Ideally, a traffic accident involving a crash would generate a large vibration impulse and could, therefore, be easily located using DFOS. This will be crucial for responding quickly, especially in remote locations where it may be less economical to install a system of hundreds of traffic cameras. Even if a traffic accident does not involve a crash, it might slow down the incoming traffic and such an incident can be quickly picked up by the DFOS. It is, therefore, important to develop a system capable of estimating the average traffic speed in real-time along the entire fiber cable.

When attempting to analyze the measured vibrations of passing vehicles using DFOS, there is a significant difference between the ideal vehicle patterns shown in Figure 1 and the real-world noisy patterns. Figure 2 illustrates a 2 min × 2 km portion of vibration data as measured by DFOS at each sampling time instant and at each sampling location along a fiber cable. Such time-location encoding of vibration data is also commonly referred to as waterfall data. When observing real-world waterfall data, it is evident that estimating traffic speeds and vehicle counts from such noisy vibration patterns is a non-trivial task. Figure 2 also lists useful qualitative interpretations like vehicle spacing, congestion, and traffic flow trends of the illustrated waterfall data. In addition to the listed interpretations, we notice inconsistencies in the measured vibration of a vehicle at different locations. This could be because of lane changes, uneven installation of fiber cables, the presence of tollbooths, and exits on highways. We also notice that each vehicle has a unique vibration pattern, which typically depends on the vehicle’s weight and dimensions.

Figure 2.

An illustration of distributed fiber-optic sensing waterfall data and some of its useful interpretations.

To develop methods for estimating quantitative attributes from DFOS data, we explored the concepts of signal and image processing. However we found that estimating traffic flow properties using unsupervised feature extraction techniques (like edge detection, optical flow, and Hough and Fourier transforms) are either ineffective or computationally non-scalable for real-time analysis of noisy DFOS data across a 45-km highway segment. This motivated us to use the concepts of deep-learning based supervised feature extraction. A disadvantage of using supervised techniques is that they require an enormous number of data samples that are balanced, diverse, and exhaustive. Creating a balanced database necessitates the understanding of various factors which influence the measured vehicle vibrations. Accumulating such a large enough waterfall database and labeling it is expensive and requires significant manpower. In this work, we developed a computationally efficient traffic monitoring system that is robust to the noises in waterfall data and does not require an expensive data collection and labeling process. The following is an overview of all the contributions presented in this work.

We proposed data pre-processing techniques: normalization and localization. Normalization overcomes the problem of varying vibration intensity gains with distance from the sensing box and the type of road structures. Localization overcomes an inability to localize observed traffic because of additional fiber cable segments present at intermediate cable monitoring centers and the junctions of roadway structural changes (e.g., from ground to tunnels or bridges).

We proposed a data generator which outputs diverse pairs of synthetic noisy waterfall data and its corresponding average traffic speed for training robust feature-extraction models.

We proposed SpeedNet, a deep neural network that is trained to estimate average traffic speeds using synthetically generated data and tested on real waterfall data.

We demonstrated the effectiveness of the above proposed methods to monitor traffic over a 44.5-km highway section using fiber cables already laid underground. Doing so differentiates our work from existing literature which either monitors short roadway sections or uses an expensive fiber-optic infrastructure strategically installed for DFOS purposes.

Challenges and Solutions of Highway Monitoring using DFOS

To explain clearly the challenges in using DFOS, we first provide the details of our highway monitoring field test in Japan. Then we detail the challenges faced when using an existing fiber-optic cable for vibration sensing and provide ways of overcoming them.

Set-up and Data Acquisition

We conducted our experiment at the Tomei expressway in the Shizuoka prefecture of Japan by collaborating with the Central NEXCO (Nippon Expressway Company) Limited who enabled us to use an optic fiber cable already laid beneath the highway for distributed sensing. An overall geographic distance of 44.5 km from Shimizu IC (interchange) as far as Numazu IC along the highway was used for our experiment. We measured the DFOS waterfall data along this sensing portion of the cable for a total duration of 2 days over a weekend. The geographic layout of our experiment, a 5-min sample of the measured waterfall data and a rough layout of the fiber cable are shown in Figure 3. Vibration amplitudes were sampled approximately every 4.08 m (spatial resolution) along the entire cable every 250 ms (time resolution).

Figure 3.

(a) Geographic portion of Tomei expressway used for our experiment setup, (b) 5-min sample of distributed fiber-optic sensing waterfall data measured from Shimizu IC to Numazu IC, and (c) rough layout of fiber cable (beneath the expressway) attached to our sensing box.

Challenges of DFOS

Uneven Intensity Gains

There are two main factors which affect the measured vibrations: distance from sensing box, and the type of road structure. The gain of our measurements decreases as we move further away from the sensing box as there are lesser amounts of light reaching the distant points of the fiber cable. Taking the ground vibration amplitude as a reference, tunnels dampen the vibrations of passing vehicles while bridges amplify them because of their own natural vibration. Such variations in the gains of vibration intensity pose a challenge when inferring traffic speeds from waterfall data.

Additional Fiber Cable Segments

There are six geographical locations that mark the sensing portion on the Tomei expressway. Each of these locations acts as a junction point where segments of fiber cables can be joined together. So, we attach our distributed vibration sensing box at the first junction, Shimizu IC, from where we record the waterfall data. The segment of fiber cable between first and second junctions, and between second and third junctions are connected at Okitsu TN (tunnel). Similar connections are made at remaining junction points and finally the cable is terminated at Numazu IC. Also, at each junction point, a few hundreds of meters of fiber cable is coiled and stored for precautionary measures. Furthermore, roadway structural changes often require additional cable segments at the junctions connecting roads to tunnels, bridges, and highway exits. We observed that approximately 48.6 km of fiber cable was used to connect the start point (Shimizu IC) and end point (Numazu IC), which is 4,100 m more than the 44.5 km of roadway connecting them. This inconsistency significantly hindered our ability to localize the observed traffic using waterfall data.

Irregular Fiber Layout

The fiber cable already laid beneath the expressway does not follow any particular section of the road consistently. For simplicity, only a rough layout is shown in Figure 3. However the actual layout in some regions changes as often as every 500 m. This means that the contribution of a vehicle’s vibrations traveling in the same lane can vary abruptly because of alterations in the fiber layout. To overcome this, one may choose to lay a new set of fiber cable(s). However, it is not economical to install 50 km of fiber cables costing up to a million U.S. dollars.

From the above observations we can summarize that the effectiveness of DFOS based traffic monitoring depends on a vehicle’s distance from the sensing box, the pavement materials it travels on, the ability to localize it accurately, and the fiber layout. In the following sub-sections we present novel solutions to overcome the major challenges of DFOS based traffic monitoring.

Overcoming Challenges of DFOS

Data Pre-Processing: Normalization

An important challenge when using waterfall data is that it is affected by vibration amplification caused by bridges and dampening by tunnels. To overcome this, we perform a column-wise normalization on the 2-dimensional matrix of waterfall data, where each column represents a time series of vibrations measured at a sensing point on the fiber cable. As discussed earlier, the gain of measured vibrations decreases as we move away from the sensing box (the start point of the sensing). Since normalization sets the sum of all values in each column to 1, the distant sensing points get the same weighting as the nearby points. The effect of normalization can be seen in Figure 4.

Figure 4.

A 5- min sample of (a) unprocessed waterfall data, (b) the same data normalized to overcome the problem of varying road type, and (c) the data corrected for localization. After the correction, the localization localization is decreased by at least 70%.

The zig-zag nature of the fiber layout is clearly evident from the normalized waterfall data in Figure 4. Vehicles traveling toward the sensing box (from right to left) are primarily measured by the last 15 km of the fiber cable, while the vehicles traveling in both directions are measured by the first 30 km. A few of the bridges, tunnels, and roads are marked in Figure 4. It is also seen that the vibration of bridges is very strong and often overshadows the vibrations of passing vehicles, thus distorting their individual tracks. However, this drawback enables the fiber cable beneath bridges to output a waterfall pattern that is uniquely characteristic of bridges.

Data Pre-Processing: Localization

After normalization, we try to improve the localization accuracy by identifying abnormal patterns in the waterfall data. A normal traffic pattern can be defined as a set of straight lines, each representing a vehicle track. However, an abnormal pattern in the form of a set of horizontal lines generated by extra segments of fiber cable being coiled and stored at some location. Other abnormal patterns are generated from a continuous fiber layout being disrupted by railway intersections and crossroads. A few such patterns are marked on the normalized data in Figure 4 and are removed when trying to localize the waterfall data. As previously mentioned, there is a difference of 4,100 m between the overall fiber length and road length in our desired sensing portion. To estimate the localization error, we utilize the unique characteristics of bridge vibrations. The normalized waterfall data of a bridge looks like a noisy rectangle, whose width equals the bridge length. We can, therefore, estimate the distance of each bridge along the fiber and along the road. Using this property of bridges, we compared the localization errors along the entire sensing portion before and after removing the abnormal traffic patterns, and shown in Figure 4. We term this process ‘localization correction’ and it decreased the difference between fiber length and road length from 4,100 m to 1,100 m. This ensures a decrease of more than 70% in the localization error at any given sensing point. Note that in this paper, the abnormal pattern and bridge pattern identification are done by manually observing the normalized data. It is also possible to develop normal vs. abnormal pattern classifiers and bridge vs. road vs. tunnel classifiers using the column vectors of waterfall data. We shall develop such classifiers as part of our future work.

Synthetic Data Generator

To maintain consistency, we first note that the training and analysis of overall waterfall data should be done in small patches of 1-min and 1-km intervals. To effectively train a deep neural network, we need at least tens (if not hundreds) of thousands of labeled training data samples. As our field test measures data for a duration of 2 days, we obtain a total of 2 × 24 × 60 = 2,880 min of waterfall data. Since the entire sensing portion spans roughly 45 km, we have 2880 × 45 ≈ 130k of 1 min–1 km waterfall patches. However to avoid any bias in our predictions, it is necessary for the speed values available for training to be uniformly distributed across the overall speed range of 10–150 km/h. This cannot be ensured as congestion on the expressway is rare and almost all the average traffic speeds typically range from 80 km/h to 100 km/h, thus making the 130k patches insufficient to train slower and much faster traffic speeds. Even if one were to measure waterfall data over several months to collect sufficient data samples below 80 km/h and above 100 km/h, labeling each sample with its average speed is expensive and time consuming.

We overcome this challenge by developing a waterfall data generator that can generate synthetic patches of training data from a given set of parameters. A list of these parameters and their accepted ranges are detailed in Table 1. Given a valid parameter configuration, the generator creates a synthetic patch of data that closely resembles the characteristics of measured real-world waterfall patches. A few samples of real and generated waterfall patches are shown in Figure 5. For each synthetic data generation, the parameters are chosen randomly from their accepted values.

Table 1.

A List of Parameters of the Synthetic Data Generator and their Accepted Values

Parameter	Choice
Average traffic speed	10–150 km/h
Number of vehicles	3–30
Number of opposite lane vehicles	0–10
Degree of noise	low, medium, or high
Thickness of vehicle tracks	low, medium, or high

Figure 5.

Two samples for each of the real-world and synthetic data used for the training of SpeedNet.

SpeedNet: Model Training and Testing

In this work, we propose a deep neural network-based speed estimation model called SpeedNet. Its purpose is to take a waterfall patch as a 2-dimensional matrix input and give its corresponding average traffic speed as output. In line with the common deep learning practices, we first train the SpeedNet model and then use it for testing as shown in Figure 6.

Figure 6.

Block diagram illustrating the training and testing processes of the SpeedNet model using distributed fiber-optic sensing waterfall data for estimation of average traffic speeds.

For the choice of architecture for SpeedNet, we used the well-known VGG16 model ( 31 )[30. Simonyan]. Briefly, the VGG16 is a neural network with 16 convolutional layers which are paired with a couple of perceptron layers and a final output neuron. Each of the convolutional layers aims to extract features from the input waterfall patch ( 32 ). The two perceptron layers aim to use the extracted waterfall features to give a value of estimated traffic speed at the output neuron. We use a ‘Linear’ activation function at the output neuron and encode the traffic speed labels by dividing them with the maximum speed (150 km/h). During testing, the value estimated at the output neuron is decoded by multiplying it by the maximum speed.

We train the SpeedNet model according to the block diagram shown in Figure 6 using only the labeled synthetic data. Each input waterfall patch is pre-processed by normalizing its columns. For this purpose, we generate 150,000 training samples and an additional 15,000 samples for validating the trained model. The speed estimation error on these validation samples gives an idea of the degree to which our model generalizes the desired task. After training the model for 100 epochs with a mean square error (MSE) loss function, SpeedNet’s training and validation errors were 97% and 93%, respectively. Weights of the trained SpeedNet model are stored for later use.

During the testing phase, we would like to estimate the average traffic speeds in measured waterfall data. This is done by first separating the waterfall data into 1-min × 1-km sized patches and then pre-processing them using normalization and localization correction. Processed patches of data are given as input to the trained SpeedNet model, which gives a speed value as output. We then perform post-processing on the output value to get the final estimated average traffic speed. We note that our model is only trained on synthetic data and is therefore not fully capable of understanding the real waterfall data noises, especially those originating from lengthy bridges and irregular fiber layout. Such noises sometimes lead to sudden jumps in estimated traffic speeds. As average traffic speeds on highways over 1 km do not drastically change every minute, we perform post-processing to smoothen our estimations and remove any outlier speed values. We choose the ‘Savitzky-Golay’ filter for our smoothing function ( 33 ). The final estimated traffic speeds using our proposed SpeedNet model are compared with nearby existing traffic camera and loop detector sensors for validating the performance of our highway monitoring system.

Results

Using the trained SpeedNet model, we estimate the traffic speeds for the entire 45 km of the fiber’s sensing portion for the duration of 2 days over a weekend. As previously discussed, the waterfall data is separated into 1-min patches each spanning 1 km. Each patch is pre-processed and the average traffic speed is then estimated. After obtaining the estimated speeds (once every minute), the time series of speed values at each sensing point is post-processed using a smoothing filter of window size 13 and polynomial order 2. To estimate speeds from measured DFOS data once every 1 min and 1 km for the entire 45-km range, an Intel 4 Core i5 processor with a 16GB RAM took roughly 40 s of computational time. It is, therefore, sufficiently fast to realize a real-time system as there is no data queued for analysis.

Comparison with Loop Detectors

To determine the performance of our DFOS-based traffic-monitoring system, we compare our estimated speeds to that of neighboring loop detectors. See Table 2. For the entire 45 km of the sensing portion, there are seven loop detectors located at various distances from the sensing box. Speeds measured at the loop detectors were provided by NEXCO for the duration of our experiment. The performance of the loop detectors themselves, as compared with the actual traffic speeds, was specified to be at least 97%. As there is significantly less traffic during the early hours, that is, from midnight to 06:00, we report the average speed estimation accuracy separately for this time interval and the remaining part of the day. The comparison metric between the DFOS and loop detector speeds is defined as:

error = \frac{\sum_{each minute} | estimated speed - loop detector speed |}{\sum_{each minute} loop coil speed} .

Table 2.

Speed Estimation Error of Distributed Fiber-Optic Sensing as Compared to Loop Detector

Loop detector location	Saturday		Sunday
Loop detector location	00:00–06:00	06:00–24:00	00:00–06:00	06:00–24:00
9.0 km	8.0%	6.6%	7.8%	7.5%
10.8 km	7.3%	7.5%	7.6%	9.3%
12.9 km	7.8%	6.4%	7.7%	6.9%
22.6 km	8.8%	7.0%	8.0%	9.3%
32.6 km	7.0%	5.4%	6.7%	7.0%
36.9 km	6.4%	4.7%	6.5%	5.8%
43.2 km	7.8%	7.1%	8.9%	9.2%

Comparison with Traffic Cameras

We also compared the traffic speeds estimated by DFOS with those from a traffic camera for an entire day. Figure 7 compares the speeds estimated by a traffic camera located 34.4 km from the sensing box, to two adjacent loop detectors located at 32.6 km and 36.9 km. Since the traffic camera is located between the two loop detectors, and assuming that the speeds on an expressway do not change significantly within a 5-km range, the traffic camera is expected to estimate traffic speeds similar to those of the two loop detectors. Figure 7 clearly indicates that the traffic speeds estimated by the two loop detectors and DFOS (in 34∼35-km patch) closely resemble each other, but the traffic camera estimations differ hugely in comparison. Figure 7 also shows 10-min DFOS waterfall data which indicate that the average traffic speed within the distance range of 32 km to 37 km does not alter significantly. This implies that in this field test, the traffic camera over-estimated the speeds by roughly 20% as compared with loop detector.

Figure 7.

Traffic speeds estimated by loop detectors and traffic cameras for an entire day as compared with speeds estimated by SpeedNet using distributed fiber-optic sensing (DFOS). A 10-min waterfall data from 05:50 to 06:00 indicates that the average traffic speed does not change significantly with a distance range of 32–37 km.

Conclusion

In this work, we present the advantages and challenges of DFOS for wide-area highway monitoring system. The overall technology is motivated by the cost-effective characteristic of DFOS in obtaining traffic information at numerous sensing points along a fiber cable. Its cost-effectiveness comes from utilizing a fiber cable that is already laid beneath most highways. At each section along the fiber cable, DFOS measures the vibration amplitude of passing vehicles every 250 ms. Since vibrations vary depending on the type of road structure (bridges, tunnels, etc.) and the fiber cable installation, they significantly affect the DFOS measurements.

We proposed novel solutions to overcome the challenges of using DFOS for traffic monitoring and developed a deep neural network (SpeedNet) to estimate the average traffic speeds every minute and every kilometer. We also proposed data pre-processing methods like normalization and localization to overcome the variations in vibration intensity gains. We further proposed a synthetic data generator for creating a labeled waterfall database for training the SpeedNet model. We demonstrated the potential of our DFOS based wide-area highway monitoring system at a 45-km section of the Tomei expressway in Japan. We also successfully demonstrated the capability of our speed estimation model, which was more than 90% accurate as compared with existing inductive loop detector sensors. As cameras suffer from weather and intensity changes, and loop detectors can be difficult to install at multiple locations, monitoring traffic using DFOS over an existing fiber-optic cable shows tremendous potential.

Our future work includes improving the SpeedNet accuracy by designing neural networks tailored for individual vehicle track detection ( 34 , 35 ). Identifying these vehicle tracks can be used as a post-processing step to fine-tune our estimated speeds. The vibration intensity and thickness of each track will be correlated to the weight and length of the vehicle. This information can be used to identify over-weight and zone-restricted vehicles. A total number of these tracks would provide a vehicle count and traffic flow rate. Finally, we can also classify these vehicle tracks to detect accidents and over-speeding, thus realizing a complete traffic monitoring system.

Footnotes

Acknowledgements

We acknowledge the cooperation of Central Nippon Expressway Company (NEXCO) Limited in providing access to the Tomei expressway and supporting the collection of data for our field trial.

Author Contributions

The authors confirm contribution to the paper as follows: study conception and design: C. Narisetty, T. Hino, M. Huang, R. Ueda, H. Sakurai, A. Tanaka, T. Otani, T. Ando; data collection: M. Huang, T. Hino, A. Tanaka; analysis and interpretation of results: C. Narisetty, R. Ueda, H. Sakurai; draft manuscript preparation: C. Narisetty, M. Huang, R. Ueda, H. Sakurai. All authors reviewed the results and approved the final version of the manuscript.

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) received no financial support for the research, authorship, and/or publication of this article.

References

Hockaday

Evaluation of Image Processing Technology for Applications in Highway Operations. 1991.

Coifman

Beymer

McLauchlan

Malik

A Real-Time Computer Vision System for Vehicle Tracking and Traffic Surveillance. Transportation Research Part C: Emerging Technologies, Vol. 6, No. 4, 1998, pp. 271–288.

Coifman

Kim

Speed Estimation and Length Based Vehicle Classification from Freeway Single-Loop Detectors. Transportation Research Part C: Emerging Technologies, Vol. 17, No. 4, 2009, pp. 349–364.

Chase

R. T.

Williams

B. M.

Rouphail

N. M.

Kim

Comparative Evaluation of Reported Speeds from Corresponding Fixed-Point and Probe-Based Detection Systems. Transportation Research Record: Journal of the Transportation Research Board, 2012. 2308: 110–119.

Matsuyama

Ukita

Real-Time Multitarget Tracking by a Cooperative Distributed Vision System. Proceedings of the IEEE, Vol. 90, No. 7, 2002, pp. 1136–1150.

Liu

Jin

Bai

Wang

Distributed Fiber-Optic Sensors for Vibration Detection. Sensors, Vol. 16, No. 8, 2016, p. 1164.

Wellbrock

G. A.

Xia

T. J.

M. F.

Huang

Chen

Salemi

Y. K.

Huang

Wang

First Field Trial of Sensing Vehicle Speed, Density, and Road Conditions by using Fiber Carrying High Speed Data. Proc., Optical Fiber Communications Conference and Exhibition, San Diego, CA, IEEE, New York, 2019, pp. 1–13.

Barrias

Casas

J. R.

Villalba

A Review of Distributed Optical Fiber Sensors for Civil Engineering Applications. Sensors, Vol. 16, No. 5, 2016, p. 748.

Hall

A. J.

Minto

Using Fibre Optic Cables to Deliver Intelligent Traffic Management in Smart Cities. Proc., International Conference on Smart Infrastructure and Construction: Driving Data-Informed Decision-Making, ICE Publishing, Churchill College, Cambridge, 2019, pp. 125–131.

10.

Akerly, E. NDDOT TRIP Program: Distributed Acoustic Sensing (DAS) along I-29 for Traffic Monitoring. National STIC Network Meeting, 2018.

11.

Danielson

B. L.

Optical Time-Domain Reflectometer Specifications and Performance Testing. Applied Optics, Vol. 24, No. 15, 1985, pp. 2313–2322.

12.

Juarez

J. C.

Taylor

H. F.

Field Test of a Distributed Fiber-Optic Intrusion Sensor System for Long Perimeters. Applied Optics, Vol. 46, No. 11, 2007, pp. 1968–1971.

13.

Haque

Rahman

Hadiuzzaman

Hossain

Siam

M. R.

Qiu

T. Z.

PARTS-Based Real-Time Vehicle Detection for Flow Measurement Considering Shadow and Illumination Variation. Presented at 96th Annual Meeting of the Transportation Research Board, Washington, D.C., 2017.

14.

Tok

Hernandez

S. V.

Ritchie

S. G.

Accurate Individual Vehicle Speeds from Single Inductive Loop Signatures. Presented at 88th Annual Meeting of the Transportation Research Board, Washington, D.C., 2009.

15.

Jeng

S. T.

Chu

A High-Definition Traffic Performance Monitoring System with the Inductive Loop Detector Signature Technology. Proc., 17th International IEEE Conference on Intelligent Transportation Systems, Qingdao, China, IEEE, New York, 2014, pp. 1820–1825.

16.

Hsieh

J. W.

S. H.

Chen

Y. S.

W. F.

Automatic Traffic Surveillance System for Vehicle Tracking and Classification. IEEE Transactions on Intelligent Transportation Systems, Vol. 7, No. 2, 2006, pp. 175–187.

17.

Bao

Dhliwayo

Heron

Webb

D. J.

Jackson

D. A.

Experimental and Theoretical Studies on a Distributed Temperature Sensor Based on Brillouin Scattering. Journal of Lightwave Technology, Vol. 13, No. 7, 1995, pp. 1340–1348.

18.

Chen

Zhang

Vehicle Classification from Traffic Surveillance Videos at a Finer Granularity. Proc., International Conference on Multimedia Modeling, Springer, Berlin, Heidelberg, 2007, pp. 772–781.

19.

Hoose

AEFFECTS: An Image Analysis Tool for Motorway Surveillance. Traffic Engineering & Control, Vol. 33, No. 3, 1992, pp. 140–147.

20.

Coifman

Revisiting the Empirical Fundamental Relationship. Transportation Research Part B: Methodological, Vol. 68, 2014, pp. 173–184.

21.

Cao

Jiang

Zhang

Guo

A Unified Framework for Vehicle Rerouting and Traffic Light Control to Reduce Traffic Congestion. IEEE Transactions on Intelligent Transportation Systems, Vol. 18, No. 7, 2016, pp. 1958–1973.

22.

Pan

Popa

I. S.

Zeitouni

Borcea

Proactive Vehicular Traffic Rerouting for Lower Travel Time. IEEE Transactions on Vehicular Technology, Vol. 62, No. 8, 2013, pp. 3551–3568.

23.

Tejedor

Macias-Guarasa

Martins

H. F.

Pastor-Graells

Corredera

Martin-Lopez

Machine Learning Methods for Pipeline Surveillance Systems Based on Distributed Acoustic Sensing: A Review. Applied Sciences, Vol. 7, No. 8, 2017, p. 841.

24.

Strong

A. P.

Lees

Hartog

A. H.

Twohig

Kader

Hilton

Mullens

Khlybov

Sanderson

An Integrated System for Pipeline Condition Monitoring. Proc., International Petroleum Technology Conference, Doha, Qatar, 2009.

25.

Peng

Duan

Rao

Y. J.

Real-Time Position and Speed Monitoring of Trains using Phase-Sensitive OTDR. IEEE Photonics Technology Letters, Vol. 26, No. 20, 2014, pp. 2055–2057.

26.

Papp

Wiesmeyr

Litzenberger

Garn

Kropatsch

A Real-Time Algorithm for Train Position Monitoring using Optical Time-Domain Reflectometry. Proc., IEEE International Conference on Intelligent Rail Transportation, Birmingham, IEEE, New York, 2016, pp. 89–93.

27.

Xie

Jiang

Zhang

Ultra Long Distance Distributed Fiber-Optic System for Intrusion Detection. In Advanced Sensor Systems and Applications V ( Wang

Fan

Bao

Culshaw

Liao

, eds.), International Society for Optics and Photonics, Beijing, China, Vol. 8561, 2012, p. 85611.

28.

López-Higuera

J. M.

Cobo

L. R.

Incera

A. Q.

Cobo

Fiber Optic Sensors in Structural Health Monitoring. Journal of Lightwave Technology, Vol. 29, No. 4, 2011, pp. 587–608.

29.

X. W.

Y. H.

Han

J. P.

Structural Health Monitoring of Civil Infrastructure using Optical Fiber Sensing Technology: A Comprehensive Review. The Scientific World Journal, Vol. 2014, 2014, pp. 1–11.

30.

Rodriguez

Casas

J. R.

Villalba

SHM by DOFS in Civil Engineering: A Review. Structural Monitoring and Maintenance, Vol. 2, No. 4, 2015, pp. 357–382.

31.

Simonyan

Zisserman

Very Deep Convolutional Networks for Large-Scale Image Recognition. arXiv Preprint arXiv:1409.1556, 2014.

32.

LeCun

Bengio

Convolutional Networks for Images, Speech, and Time Series. The Handbook of Brain Theory and Neural Networks, Vol. 3361, No. 10, 1995, p. 1995.

33.

Savitzky

Golay

M. J.

Smoothing and Differentiation of Data by Simplified Least Squares Procedures. Analytical Chemistry, Vol. 36, No. 8, 1964, pp. 1627–1639.

34.

Gkioxari

Dollár

Girshick

Mask R-Cnn. Proc., IEEE International Conference on Computer Vision, Venice, Italy, IEEE, New York, 2017, pp. 2961–2969.

35.

Zhang

Ren

Sun

Deep Residual Learning for Image Recognition. Proc., IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, NV, IEEE, New York, 2016, pp. 770–778.