Applications of fiber optic sensors in traffic monitoring: a review

Instrumenting pavement with fiber optic sensors has recently gained popularity as a part of the digital infrastructure transformation. In this survey, we present some of the recent real-world implementations of fiber optic sensors as real-time traffic monitoring systems. We highlight advantages and limitations of the surveyed instrumentations. We identify preferred range of values for some of the key design parameters that are necessary for durability and accurate weighing of individual axles. These parameters include strain resolution, scanning rate, installation depth, spatial resolution, and sensor orientation. Some of the studies report excellent results in terms of vehicle axle count, weight, and speed achieving accuracy levels above 90% with single digit margin of error. Lastly, we draw attention to some of the outstanding challenges from technical, operational, and scalability points of views.


Introduction
Traditional traffic data collection is typically carried out by a group of people for a specific site. The findings are documented in terms of traffic flow, capacity analysis, and traffic prediction models. This process is repeated after some time, typically several years. This type of method has several drawbacks as the data are collected at a periodic fashion and may not represent the traffic behavior accurately. Technologies utilizing video cameras, radars, magnetic fields, and wireless sensors are proposed and implemented at some sites to collect real-time data, improving the traditional methods. However, these methods are expensive and difficult to install and maintain. They are also easily damaged and influenced by the environment. In order to overcome the aforementioned difficulties of existing traffic monitoring methods; robust, cost-efficient, accurate, consistent, and safe solutions are needed.
Next generation road infrastructure will host various sensing and communication elements. Connected electric autonomous vehicles (CEAVs) will charge their batteries through wireless power transfer from the pavement. Furthermore, CEAVs will communicate with the roadway through wireless communications. This will add another layer of safety for autonomous vehicles. Traffic monitoring sensors will be an integral part of the next generation road infrastructure. Fiber optic sensors (FOSs) have already been installed in various ways as traffic monitoring systems. Majority of the reviewed installations utilize FOSs to measure the strain and vibration passing vehicles induce. It was shown that the systems were able to collect vehicle count, classification, axle count, speed, and weight. Monitoring and managing the real-time traffic flow can help in adjusting traffic signal control and traffic regulation, reducing traffic congestion, and detecting atypical behavioral events [1]. This type of data gathering will also be useful in developing new pavement deterioration models and help the cities and department of transportations (DOTs) with their road maintenance efforts. There are many more applications that real-time traffic data collection will help prosper.
In this survey, we highlight some of the implementations in practice with their advantages and limitations, identify essential design parameter values for a successful system, and provide a discussion about future directions.
The remainder of this paper is organized as follows: In Sect. 2, fundamentals of FOSs are presented. Section 3 reviews various real-world implementations of FOSs. Analysis of the outcomes is given in Sect. 4. Section 5 concludes the survey with future directions.

Fiber optic sensors
Over the last few decades FOSs have gained increasing attention from researchers and engineers in various fields. They are accurate, durable, immune to electromagnetic interference, require relatively less power, and have relatively low lifetime cost [2][3][4]. Due to these virtues, they are outperforming the conventional electrical sensors such as piezoelectric sensors and strain gauges. Understanding the physical characteristics of light waves and their interactions with surroundings have led to designs that act as strain, temperature, acoustic, magnetic fields, acceleration, rotation, pressure, humidity, and viscosity sensors. They are employed in a wide range of industries such as oil and gas, energy, biomedical, civil, aerospace, and transportation [5,6].
Basic principle of all FOSs is to detect variations in intensity, phase, frequency, and polarization of light waves induced by the measured variable. Two key performance metrics of FOSs are sensitivity and resolution. Sensitivity represents the relationship between the variation of the sensor output and the corresponding variation in measured variable. A good sensor should have a high sensitivity, i.e., small variations in the measured variable should correspond to large variations in the output. Resolution describes the sensor's ability to detect small variations. It is defined as "the variation in the value of the measured variable that causes a variation in the output value equal to the uncertainty of the output itself" [5]. In other words, any variations less than the resolution value cannot be accurately sensed.
FOSs can be categorized into three types: single-point, quasi-distributed (multiplexed), and distributed sensors (see Fig. 1). Quasi-distributed sensors are essentially point sensors that are multiplexed and located on pre-determined locations.

Single-point sensors
The most straightforward single-point sensors are based on interferometry. Interferometer sensors, as the name suggests, exploit the interference between light beams. In an interferometer, the light is separated into two beams where one beam is subjected to the sensing environment experiencing a phase shift, while the other is used as a reference preserved from the sensing environment [7]. The beams then are recombined. This results in either constructive or destructive interference. Monitoring the recombined beam yields information about the measured variable. Fabry-Perot, Michelson, Mach-Zender, and Sagnac interferometers are some of the well-known configurations. A comprehensive review of interferometric sensors is presented in [8,9].
There are also grating based single-point sensors such as fiber Bragg grating (FBG) and long period grating. A typical FBG has a sub-micron period whereas a long period grating has a period in the range 100 m to 1 mm [10]. A thorough review of long period grating sensors is given in [10,11].
FBG based sensors occupy a large market share due to a number of advantages: small form factor, lightweight, selfreferencing, easy to multiplex, no need for electrical connections and the compatibility for non-invasive remote sensing [5]. Self-referencing is a very important strength because the output does not directly depend on the total light intensity or losses in the fibers and couplers. The output of an FBG sensor is given in terms of an absolute parameter (wavelength) [8]. This property makes FBG sensors more robust, accurate, and relatively simple to operate.
The basic working principle of an FBG is illustrated in Fig. 2.
A broad-spectrum light beam is transmitted through a fiber cable with an FBG. The FBG reflects a subset of the broad-spectrum light back to the source and transmits all other light. The resulting back-propagating spectrum presents a peak centered at the Bragg wavelength. The spectrum of the beam that passed through resembles the spectrum of a notch filter. The Bragg wavelength is dependent on the grating pitch, which is a design parameter, and the refractive index of the fiber. The reflected Bragg wavelength, B , is defined as: where n eff and Λ denote the effective refractive index and the grating pitch (periodic spacing of the grating), respectively. A variation in the environmental parameters, for instance, temperature and strain, influence both the pitch ( Λ ), and the refractive index ( n eff ). This causes a shift in the reflected spectrum (see Fig. 3). In sensing applications, the Bragg wavelength (central frequency of the reflected spectrum) is monitored.
Strain is a normalized measure of deformation, which represents the displacement particles in the matter with respect to a reference length. From a physical point of view, the strain can be decomposed into a normal and a shear component. Stretch or compression along the fibers is represented by the normal component. The shear component is related to the sliding of layers over each other. If the length of body increases, the normal strain is called tensile strain, if it reduces, the strain is called compressive strain. The strain has a dimensionless unit, and it is given as: where ΔL and L denote the change in length and initial length, respectively.
The grating reflects some of the incident light beam back. The plots represent the spectrum of the incident, reflected, and transmitted beams Applied strain causes variation in the effective refractive index and variation in the pitch. Assuming that the material properties of the grating is known, the shift in the Bragg wavelength can be given as: where p e denotes the photo-elastic coefficient and its value is 0.21 for a conventional single mode fiber SMF28 [5]. Typically, a Bragg wavelength written into a single mode fiber SMF28 will shift approximately 1.2 pm/ [5]. Temperature change also causes variation in the effective refractive index and the grating pitch due to thermal expansion. A typical temperature induced Bragg wavelength shift is approximately 13.7 pm/ • C. Therefore, it is important to consider the temperature dependence of the sensor output and devise methods to compensate for it.
The accuracy of measurement and the number of measurements per second are dependent on the optical interrogation unit. The interrogator provides the broadband light source and also tracks the reflected peak wavelengths in real time. Interrogator wavelength resolution can be as low as 1 pm, which equates to a 0.83 measurement resolution, at acquisition frequencies up to 1000 Hz [12]. The resolution reduces to around 3 pm at 5000 Hz acquisition frequency.

Distributed sensors
Distributed sensors have the ability to sense at any point along the fiber cable through scattering. When a light beam is transmitted into a fiber optic cable it experiences random Rayleigh, Brillouin and Raman scattering along the way and some of the scattered signal travel backwards. These scatterings arise from the interaction of the light waves with the particles within the fiber and with fiber glass. Fluctuations in the density of the fiber glass and thermally generated acoustic waves change the properties of these backscattered waves. The local changes in temperature, strain, and vibration along the cable will modulate the signal. Monitoring the changes in the backscattered signals yields information about the measured variable. The backscattered signals are mapped to locations along the cable depending on their arrival time to the origin. The smallest length within which a significant change in the measured variable can be detected is defined as the spatial resolution [5]. The concept of a distributed sensor is shown in Fig. 4. Acquisition frequency and accuracy vary between the different technologies, and will also depend on the length of the fiber, desired spatial resolution, and the time taken to make the measurement.
Physics of distributed fiber optic sensors are presented in great detail in [13,14].

Implementations
FOSs have been utilized in various experiments dating back to late twentieth century to demonstrate their excellent potential for traffic monitoring. Mimbela et al. [15] provide a comprehensive review of some of these earlier projects that laid the foundation for the next generation implementations. Here, we present some of the recent developments in the field.

Concrete bridge in Northern Ireland
Lydon et al. [16] equipped a reinforced concrete bridge in Loughbrickland, Northern Ireland with FBG sensors (10 mm length) to monitor deterioration of the structure, weigh individual vehicles and detect overweight vehicles (see Fig. 5).
The authors suggest that placing the sensors underneath the bridge is a better solution then placing them on the road pavement in terms of longevity and portability of the system. Their finite element analysis (FEA) showed that an overloaded 6 axle truck produced 30 , a light goods vehicle produced about 5 and a compact car on the slab soffit/ deck produced about 1 . After running trials, they found out that slab soffit strain values decrease if the vehicles travel directly over the girder. To address this problem, they placed FBGs on the girders in addition to the soffit. The consequent improvement to the axle detection accuracy is presented in Fig. 6.
One can see that the signal from the girder compared to the soffit has more distinguishable peaks.
Another challenge they faced was to distinguish individual vehicles when multiple vehicles travel side by side on different lanes [18]. They included more sensors in the transverse direction and developed post processing algorithms to overcome this. The system also included a separate FBG to compensate for temperature strain.
The axle detection accuracy of the system was 96% with ± 3.7% standard error (SE). They compared the detected axle weight estimations with the values obtained from a nearby static weighing station. Average individual axle weighing error was − 2.25% and average gross vehicle weighing error was − 4.5%. The margin of error (MOE) was 9.33% with 90.1% confidence level for the axle weighing error and 2.45% with 85.5% confidence level for the gross vehicle weighing error. The authors think that the regular under-weighing might be due to a signal processing error [19]. Calibration with different vehicle types (as opposed to one) at different speeds is anticipated to enhance the accuracy as well.
During their investigation, the authors also learnt that FBG sensors perform better than conventional strip sensors, electrical resistance sensors, and vibrating wire sensors due to various reasons. Temperature effect seems to be the most concerning problem for conventional strip sensors and electrical resistance sensors. Moreover, in order to install conventional strip sensors, one needs to destruct the road pavement. Vibrating wire sensor produces accurate results at a 0.2 Hz scanning rate. This is a very low scanning rate for live traffic data acquisition. The minimum scanning rate should be around 500-1000 Hz.

Steel girder bridge in LaGrange, Georgia
In a recent work, Oskoui [20] presented a weigh-inmotion (WIM) system on a four-span steel girder bridge in LaGrange, Georgia (see Fig. 7).
The goal of the system is to weigh vehicles, estimate axle spacing and speed. The proposed method requires twowheel sensors for axle spacing and speed calculations and a rotation sensor for axle weight calculations. They employ FBG sensors with 75 mm gauge lengths. The wheel sensors were placed directly under the deck at a transverse position where a wheel is likely to go directly over. The FBG sensor that is meant to measure the rotation at the bridge abutment was installed on a cantilever beam as shown in Fig 8. Temperature strain is compensated with a dedicated sensor that is placed on the arm of the cantilever beam with an open end. Bridge installation positions of the sensors are shown in Fig. 9.
The abutment rotation, hence strain, values are calibrated using 4 trucks with known axle weight and spacing. The estimated vehicle weights are compared with the values obtained from a nearby static weighing station. Out of 10 field trials, absolute individual axle weighing error values ranged from 0.23% to 16.67%. The average error − 3.3% ± 3.15%. Absolute gross vehicle weight (GVW) estimation error ranged from 0.06% to 12.08%. The average error was − 3.38% ± 2.88%. Axle spacing estimations were within 3% error. The average error was 0.54% ± 0.47%. Their system was not able to distinguish individual tandem axles. More advanced signal processing is needed.

Longweigang bridge in Jingdezhen, China
Long-gauge FBG sensors are utilized on the Longweigang Bridge as a part of a WIM system. The Longweigang Bridge is a simply supported reinforced concrete plate girder    [20] bridge. Chen et al. [21] argue that long gauge FBGs are more immune to cracks and damages on a structure compared to point sensors. They installed four long-gauge (0.5 m gauge length) FBG sensors on the plate girder of a bridge in Jingdezhen, China as shown in Fig. 10. The distance between each sensor was 2 m.
The calibration was carried out by a heavy truck with a gross weight of 24,929 kg and three axles. The validation trials were done with 5 vehicles; one SUV, one bus and three trucks. The maximum error values were 4.67%, 5.27% and 7.32% for velocity, wheelbase, and GVW, respectively. The average errors were 1.75% ± 3.04%, 2.03% ± 3.19%, and 1.11% ± 5.49% for velocity, wheelbase, and GVW, respectively. The errors were a bit higher than their theoretical results. They think this might be because of the higher noise level in an actual environment. The main limitation of this design was that any wheelbase smaller than the 2.5 m could not be identified due to the larger gauge length.

Concrete bridge in Sacramento, California
This project included a prestressed box-girder bridge in Sacramento, California and a set of FBG rosettes. Bao et al. [22] focused on capturing the shear strain near the abutment and comparing it to traditional flexural strain. The installed rosette sensor was made of two strain FBGs (75 mm gauge length) placed at an angle. Since the shear strain is subtraction of the strain value from one sensor from other, temperature strains cancel each other out, yielding a self temperature-compensating shear strain sensor. Both of the sensors were installed inside the concrete box-girder. The bridge and the placements of the two rosettes are shown in Fig. 11.
The calibration tests were carried away at nighttime with a dump truck with a trailer. The superiority of the shear response to the flexural response in terms of distinguishing individual axles is demonstrated in Fig. 12.
This is because the flexural strain is the result of overall vehicle weight whereas shear strain is more sensitive  to individual axle weights. Even though shear rosettes can distinguish more axles, they also fail when the wheelbase is very small (tandem axle) as seen in the figure. Absolute axle weighing errors ranged from 0.4% to 18.5%. The average error was 1.24% ± 4.76%. Measured axle spacing errors ranged from − 2.1% to 11.9%. The average error was − 1.86% ± 5.7%. Maximum and average GVW errors were 1.5% and 0.5% ± 1.2%, respectively, out of 3 samples. Similar tests were also carried out in Elmhurst Illinois and Chicago, Illinois [22].

Hampden bridge in New South Wales, Australia
The Hampden Bridge is a heritage listed bridge located in New South Wales, Australia. It is a steel suspension bridge that runs across the Kangaroo River (see Fig. 13). In 2007, the local traffic authority suspected that vehicles heavier than the bridge's load limit were using the bridge following an observed increase in unscheduled maintenance requirements. To understand whether overweight vehicles were using the bridge, Monitor Optics Systems installed FBG sensors onto key suspension rod members [23].
The FBG sensors were attached to two key suspension rod members that connected to the bridge's hinge joint in the center. This allowed a linear strain to load relationship, and thus creating a WIM bridge. A truck that weighed 42,500 kg (the bridge's load limit) was used to calibrate the sensors, driving over the bridge from each direction 3 times (total of 6 tests). The strain recorded for each test was approximately 350 , with a total of 5% variation in the strain over the 6 tests. The recorded strain for a test is given in Fig. 14.
Following a successful trial period, cameras were then implemented onto the bridge to capture the details of the offending vehicles. The strain data and vehicle pictures were sent to an online database that the local traffic authority could monitor.

Bridge prototype
Alamandala et al. [3] built a 250 cm long and 30 cm wide bridge prototype and equipped it with two FBG sensors. The prototype has five beams and six pairs of 12 cm high piers. A robotic vehicle is used in their tests. Their report indicates that there is a linear relationship with roughly 0.9 linear coefficient between the weight of the vehicle and shift of the Bragg wavelength. The maximum individual axle weight estimation error is roughly 25% out of 15 samples.

Road surface in Riga, Latvia
Grakovski et al. [24,25] placed fiber optic cables by Sensor Line in 30 mm wide gaps on the surface of the road in Riga, Latvia. The gaps were filled with resilient rubber afterward to complete the installation. This is illustrated in Fig. 15.
They used a loaded truck with speed ranging from 10 to 90 km/h. The field trials showed that the wheel size, wheel pressure and speed changed the force applied to the ground due to inertia properties. They estimated the wheel footprint (area of the wheel contact), speed, and axle weights. For typical wheelbase values, the method achieved accuracy within 4% for footprint. For tandem axles, the method needs to be improved as the error was 31% for one of the footprint estimations. For speeds above 48 km/h, the axle weighing errors were typically within 10%. However, at lower speeds, errors Fig. 13 Hampden Bridge in New South Wales, Australia [23] were in the range of 10%-20%. The authors believe that the effects of acceleration and deceleration became more apparent in the signals at lower speeds. Specifically, the vertical vibration of the vehicle was attributed as the reason for this. To mitigate this, they suggest using a platform to increase the exposure time of the load to the sensors. Overall, the average axle weighing error is − 1.81% ± 1.78%.

Concrete road in Otsego, Minnesota
Al-Tarawneh [26] presented a proof of concept for traffic monitoring using FBGs at a research facility in Otsego, Minnesota. The experiments were carried out at section Cell 40 of the test road. The site is shown in Fig. 16.
A 3-D FBG sensor (see Fig. 17   is believed to be due to the transverse position of the wheel. Vertical and transverse sensors required the wheel to be (almost) directly over the sensor whereas the longitudinal sensor was not as sensitive to that.
Al-Tarawneh also theoretically showed that in a nonreinforced slab of concrete, the sensitivity was maximum toward the top (surface of the slab) and bottom in terms of depth. The sensitivity was minimum in the middle portion of the slab. Moreover, it was found that smaller slab surface area leads to better sensitivity for the sensors. This is especially true when the slab size is smaller than 1.83 m (6 feet). Increasing the slab length beyond 1.83 m (6 feet) has an insignificant effect on the sensor sensitivity. To achieve 80% WIM accuracy with the longitudinal sensor, the center of the wheel should be within 15.24 cm (6 inches) around the sensor. Based on an analytical model for modulus elasticity of the concrete pavement, it was shown that after 10 years, the sensitivity of the sensor will drop by 10% which might introduce measurement errors.

Asphalt road in Rotterdam, Netherlands
In Rotterdam, Karabacak et al. [27] laid down two FBG equipped fiber optic cables that run perpendicular to the direction of travel and are 1.5 m apart. The cables are 15.24 cm (6 inches) deep and each host an array of FBGs that are 7.62 cm (3 inches) apart. This is illustrated in Fig. 18. They repeated the experiments at three different locations. Typical installation takes about 2-3 h of road closure.
They performed tests with trucks and compact cars with speeds up to 90 km/h. A compact car with approximately 300 kg (660 lbs) per wheel load, generated 4 . Local pavement structural variables, asphalt temperature, speed, and transverse wheel position are found to be key variables that affect the performance of the system. To mitigate undesirable effects by structural variables, temperature, and speed, a detailed calibration is needed. Temperature also needs to be tracked independently. The problem they faced regarding the wheel position was because of the gaps between the Fig. 17 The 3-D FBG sensor [26]. Units are in inches, R=radius

Fig. 18
Fiber installation in 15 mm slits in Rotterdam, Netherlands [27] sensors. If the wheel was in between sensors rather than directly on top, the weight estimations were different. Therefore, calibrations could not be done precisely and that caused inconsistent estimations. Signal processing algorithms were devised to address this. To assess the performance of their algorithm, they compare the estimated position to a theoretical model. This comparison is illustrated in Fig. 19.
This also allows the design to calculate the track width of the vehicles. Overall performance of the WIM was C-class according to the COST323 model. They reported 11% MOE for individual axle weighing. The sensitivity was high enough that even a bicycle wheel was identified.

Asphalt road in Ostrava, Czech Republic
This project employs FBG sensors fixed into the road as well. Fajkus et al. [28] chose to house the sensors in polydimethylsiloxane (PDMS), a polymeric two-component potting material due to its flexibility and strength. A 3-D printer was used for the potting process. A potted sensor and its asphalt implementation are given in Fig. 20.
The sensor was placed at a depth of 15 mm. The groove was then filled with cold asphalt. During the curing process, it was found that the Bragg wavelength increased but did not impair the functionality of the sensor. Based on the sensitivity analysis, it was established that 5 embedded sensors in a chain would cover the entire lane with a width of 3.2 m.
A couple of fiber optic cables (with 5 FBGs each) were installed in a road in Ostrava, Czech Republic 2 m apart. Live traffic measurements were taken for 18 days. The system's vehicle count was compared to an on-site camera detection system. The embedded sensors identified 3,963 vehicles missing only 16 vehicles (99.62% ± 0.19%). This a notable accuracy level. We believe that the design's high sensitivity (a car wheel inducing a 20 pm Bragg wavelength shift) is the main factor for this. The estimated vehicle speeds were compared to the reference speeds obtained by optical gates. Axle speeds were averaged to estimate the vehicle speed. Out of 3,963 vehicles, the measured speed values ranged from 38 km/h to 64 km/h. The maximum absolute speed estimation error was 5.83%. The average error was 2.65%. Fig. 19 Measured vs theoretical wheel position [27] Fig. 20 Potted FBG and in-pavement installation [28] Portable platform in Brussels, Belgium Yuksel et al. [29] designed a portable platform instrumented with FBGs (3 mm long) to monitor traffic. Typical installation is shown in Fig. 21.
There is a total of 6 FBGs where 4 of them are hosted in between the metallic plates and the other 2 are out of the platform. The extra two FBGs are used for temperature compensation. All of the sensors reside in a protective tube. The height of the platform is approximately 3 cm.
They tested their proof of concept on a road with live traffic in Brussels. In their preliminary experiments, more than 50 cars were detected in 4 h. Their implementation performs with an average error of 6.71% ± 0.73% for velocity estimations. It is inherently challenging for the design to estimate speeds as the wheelbase information needs to be known a priori. This is still work in progress and further testing and enhancements are needed to quantitative results.

Portable dynamometer
Belitsky et al. [30] propose to use a fiber optics dynamometer based on photoelasticity to identify wheel presence. The dynamometer they use is essentially made of two plates glued to each other and a single-mode optical fiber that is looped in between those plates. The plates are made of two layers; a metallic external and a rubber internal layer. They used a plastic covering for the cable. The overall thickness of the plates is approximately 1 cm. An example installation where the plates are fastened to the ground is shown in Fig. 22.
The external vertical force on the plate modulates the light and the number of periods of the output signal changes according to the absolute value of the force. The scale's resolution is roughly about 0.33% of the vehicle's weight. To reduce the errors caused by the wheel impact on the plates (as the plates are 1 cm above the pavement surface), they recommend installing the plates on speed bumps. They calculate that for a vehicle with speed 50 km/h, the measurement deviation is less than 3%. A temperature compensation is suggested as it affects the photoelasticity of the fiber optic material. Laboratory experimentation showed promising results (as low as 3% average error with 0.54% SE).

Test road in the UK
Hall and Minto [31] propose to utilize the existing spare roadside fiber optic cables as distributed acoustic sensors with minimal overhead. Their preliminary tests indicate that cars can be detected up to 30 m offset from the road. Although the best results are obtained when the offset distances are shorter than 10 m. The recommended burial depth for cables is between 10 cm and 50 cm. The spatial resolution of their distributed sensor is 10 m. They suggest two separate roadside cables on each side for divided roadways (see Fig. 23).
Authors argue that this type of roadside installation might capture traffic parameters such as average speed, queue/congestion detection and journey times. However, it would be challenging to identify individual vehicles in specific lanes with roadside installation. They suggest routing cables perpendicular to the direction of travel to collect vehicle specific statistics.

Iron mine in Ma'anshan, China
Similar to [31], Liu et al. [1] also propose to use distributed acoustic sensing to monitor the vibrations from the vehicles.  They carried out their field studies in the Ma'anshan iron mine in China. The fiber optic cables were installed on the shoulder of the road as shown in Fig. 24 covering 100 m. Cement mortar was used to fix the optical fiber to the ground. The seismic signals induced by ore trucks and pickup trucks were recorded.
The space resolution was 1 m, therefore there were 100 virtual sensors across the 100 m span. The scanning rate was 2500 Hz. They collected data for the duration of two days using loaded trucks. The results were compared to an on-site video camera. Authors developed signal processing algorithms based on Wavelet theory to increase the signalto-noise ratio (SNR) of the signals. The method achieves average vehicle detection accuracy of 86.11% ± 11.3% vehicle detection accuracy with a sample size of 36. The vehicle count detection method becomes inaccurate when multiple vehicles are passing in series close to one another. Regarding the velocity calculation, their method stays within 6% error range.

Sensitivity
Perhaps the common challenge of the presented implementations was achieving a good coupling between the monitored environment and the sensor. For bridge instrumentation, the location (girder, soffit, support, abutment) and orientation of the sensors were key factors in producing a consistent Fig. 23 Test road roadside fiber cable deployment in the UK [31] Fig. 24 Iron mine roadside cable installation in Ma'anshan, China [1] detection method. Making those choices required extensive field trials and/or FEA since every bridge is structurally different.
From an embedded in-pavement sensor installation perspective, the depth and orientation of the sensors played significant roles. For slabs of concrete, depths near the surface and near the bottom achieved the highest coupling [26]. Longitudinal sensors had the largest radius of influence compared to vertical and transverse ones. This is because they produce high enough strain signal peaks regardless of the exact transverse position of the tire. Even though transverse and vertical sensors in some cases produce higher strain signal peaks than the longitudinal sensors when tire is exactly on top of the sensor, the strain values drastically decrease when there is offset between the tire and sensor positions. Therefore, an array of longitudinal sensors can cover a given area with the least number of sensors making them a good practical choice in terms of cost. Regarding the distributed roadside installations, the highest coupling factor was reached when the cable was placed on the road shoulder.
For asphaltic pavement, closer the sensor is to the surface, generally higher the sensitivity becomes. If the sensors are too close to the surface, this may cause durability problems. The sensors can be placed in specific housing such as or structures to amplify the strain values. For example, authors in [28] encapsulated FBGs with polydimethylsiloxane to provide protection and high sensitivity. Authors in [16] adhered FBGs on to a metal plate to amplify the strain values. It is important not to amplify the noise floor in the signal when doing so.

Resolution
To detect individual axles, high spatial and temporal resolutions are needed. The spatial resolution of the sensing elements should be approximately maximum 50 cm. This was one of the disadvantages of long-gauge FBGs and some of distributed sensors. Sampling frequency of the sensors should be minimum 500-1000 Hz. Embedded and bridge implementations required sensor resolution of 1 to identify a typical passenger car. Systems operating at frequencies above a kilohertz with such resolution generally require expensive equipment. Advanced signal processing techniques can be developed as shown in [32] to improve sensor resolution values with relatively cheaper equipment.
FBG sensors are excellent choices for traffic monitoring because they can operate at very high rates with excellent spatial and sensor resolution. A typical FBG with spatial resolution of 1 cm can operate at 5000 Hz with 1 pm resolution. However, distributed sensors typically have a subset of those features. For instance, the distributed sensor in [1] has 2500 Hz scanning rate but with only 1 m spatial resolution. 1 m spatial resolution might be not enough to distinguish individual axles of a heavy weight trucks. The distributed sensor mentioned in [13] has a spatial resolution of 2 cm but with sensor resolution of 35 . There are distributed sensors given in [33,34] with satisfactory resolutions but with scanning rates that are not suitable for traffic monitoring.

Durability
Protection of the fiber optic sensors is an ongoing challenging issue to deal with [35]. Fiber optic cables need appropriate packaging to survive harsh conditions. The packaging material needs extensive consideration as the selection can be a trade-off with the sensitivity of the sensor. Glass fiber reinforced polymer, polydimethylsiloxane, rubber, and plastic were among the chosen materials for robustness. In addition to sensor protection, the installation approach and the structure deterioration are important. A case of asphalt fiber installation based on micro-trenching (5-10 cm deep) was shown to crumble in about a year [36]. In general, it is challenging to evaluate the durability of the systems since they need to stay alive for a few decades and be monitored throughout. The surveyed installations were proof of concept type projects and were active for a maximum of one year.
A roadway section is typically designed for 1 million equivalent single axle loads (ESALs) per year. Therefore any permanent roadway sensor installation should last the life of a typical roadway with a minimum of 15 million ESALs and a target of 50 million plus ESALs. ESAL is a metric that is used by the pavement engineers to refer to the effects of axle loads on the pavement. By definition, 1 ESAL is equal to a 8165 kg (18,000 lbs) axle [37]. For example, an ESAL value of 4 is equal to the effect on the pavement a 8165 kg axle would cause after 4 passings. Note that ESAL and weight of the axle load has roughly a fourth-power relationship. The pavement is designed according to the expected annual ESAL criteria of the section it is incorporated into. For example, pavement depth and pavement construction materials are modified depending on the expected ESALs to meet the durability requirement.
After the roadway section is installed, the real-time ESAL count should be calculated using the monitored traffic data. The real-time ESAL count may indirectly provide information about the deterioration status of the roadway. Using the live deterioration indicator, the maintenance planning can be achieved proactively. For example, if a roadway detects a 15% higher ESAL rate than the roadway was initially designed for, it will likely need maintenance several months before the initially anticipated time. In a study on a federal highway (BR-381) in Brazil, it was found that the accumulated ESAL was 6 million in 37 months (Dec. 2014 -Dec. 2017) as opposed to the roadway's design ESAL of 5 million [38].

Signal processing
In some cases, the induced strain values by the vehicles were not significantly larger than the resolution of the sensors. This may be the case when there is not a satisfactory coupling between the sensors and the induced strain. In these cases, advanced signal processing techniques might be necessary to improve the signal to noise ratio. Several of the studies showed that signal processing improved the ability to detect axles. Adaptive filtering, Wavelet theory, and artificial neural networks were some of the tools utilized. Wavelet denoising [39] was especially useful in denoising and capturing time-frequency response of the signals better than conventional methods. Even though it is true that signal processing is helpful in achieving better results, it does bring an additional computation layer. One needs to be careful in sticking to simple enough methods as the data needs to be processed in real time.

Performance
Fiber optic instrumentations have demonstrated remarkable results with accuracy performances above 80% and in some cases above 90% with a single digit MOE in terms of vehicle counts. Axle weighing and wheelbase calculation errors were generally within 5%-10%. Some of the major problems faced were temperature induced strains in the sensors, vehicles traveling in neighboring lanes, and vehicles with tandem axles.
Even though some of the reported performance metrics are excellent, the number of measurements taken are limited to a few tens of vehicles in most cases. According to European specifications for WIM, COST 323 model, it is suggested that at least several thousand vehicles should be tested to measure the performance of the system [40]. The calibration runs are suggested to be carried out with vehicles that are representative of the expected traffic. The margin of error should be ± 15% and ± 20% for GVW and axle weights, respectively. These accuracy levels refer to COST323 B+ accuracy class and better [41]. Axle weighing is a complex problem as there are variety of factors contributing to the accuracy of the measurements. Some of these are grade, slope, and curvature of the roadway; speed, suspension system, tires, and aerodynamic characteristics of the vehicle; and environmental factors such as the wind and temperature [15,42].

Summary
In Table 1, we provide a summary of the key findings of each implementation.
In light of the surveyed projects, we identified preferred design parameters of a traffic monitoring system using fiber optic sensors. Please see Table 2.

Conclusions and future directions
Real-world projects have demonstrated the capability of FOSs to count axles, weigh axles and estimate speed of the vehicles. FOSs were instrumented in cementitious concrete, asphaltic concrete, road shoulders and on bridges. Calibrating the measurements and compensating for temperature caused strains were vital in producing accurate estimations. Design parameters such as spatial resolution, scanning rate and strain resolution were found to be vital for a precise system. Each project had its own site-specific challenges to overcome in terms of ease of installation, achieving a good coupling between the structure and the sensor, and sensor protection.
There are some technical outstanding issues that need addressing. Typically, studies assume that every vehicle is made of two axles. If not, they group the spikes (axles) in the strain signal if they are detected shortly after another. This inherently assumes that the speed of the vehicle is consistent across the measurement area which does not match the readily observed practices of vehicles in a real-world environment. What is more, traffic can come to a full stop and erroneous axles can be grouped together. To the best of our knowledge the axle grouping issue has not been addressed yet. This problem can be attempted to solve by having a very dense sensing platform and an intelligence that is capable of tracking all axles at all times. Neural network-based intelligence is a good candidate for this purpose. We believe that such a solution would integrate well with the upcoming connected electric autonomous vehicles providing a fully sensitive network of roads.
Vehicle dimensions such as wheelbase and track width may also be collected. Such information may be combined with vehicle weight and axle count to classify vehicles. Due to MOE in estimations and inescapable variables such as driver weight, number of passengers, and amount of extra load in vehicles, classification is more likely dependent on a probabilistic model. This information is useful in understanding the traffic behavior and city planning. One such categorization standard is given by Federal Highway Administration [45].
Embedded traffic monitoring systems are expected to be a part of the next generation digital infrastructure. To that end, we should start thinking about operational functionalities and scalability of the proposed solutions. In other words, we should ask the question: What is the best means of deploying FOS based traffic monitoring systems?  Sensor type FBG FBG based systems can have a combination of high spatial resolution, scanning rate, and strain resolution. Typically, distributed fiber optic sensing systems need to make a trade-off between these parameters [13] Spatial resolution x ≤ 50 cm Long-gauge FBG installations showed that spatial resolution of larger than 50 cm will likely result in failed individual axle detection or tire, especially for tandem axles [20,21] Scanning rate x ≥ 500 Hz To correctly capture the strain signal peaks, the sampling rate of the sensing system should be approximately at least 500 Hz [16] Strain resolution x ≤ 1 In concrete bridge and pavement instrumentations, strain resolution of 1 was necessary to detect light-weight vehicles [16,26] Sensor orientation Longitudinal Embedded sensors that were aligned with the direction of travel exhibited larger radius of influence compared to vertical and transverse ones. [26,43,44] Sensor depth x ≥ 10 cm Surface and near surface installations do not last long relative to the lifetime of the pavement [36]. Embedding sensors encased in a protection material at least 10-15 cm deep from the pavement surface is preferable Instrumentation Embed in pavement Road shoulder implementations struggle with identifying and weighing individual vehicles [1,31]. Bridge installations are not applicable to roadways For such systems to be widely adopted, they should be easy to install, upgrade or replace, be consistent and be userfriendly. For example, portable solutions might be a good option for certain private avenues but are probably not a viable one for public road adoption as they cause elevation in the road pavement surface by a few centimeters. Durability would be a colossal issue as it would directly be impacted by passing vehicles. It might also cause problems for snowplow trucks and vice versa. Bridge sensing systems are tough to scale as every site would require a detailed FEA and extensive calibration.
Embedding sensors within cementitious concrete pavement at 15 cm or higher depth appears to be the most favorable method in terms of durability. Having the sensors within the pavement means they can be used on bridges and tunnels as well. To help with the consistency, standardized precast pavement slabs embedded with sensors may be a practical choice. From operational perspective, this may reduce installation times, standardize delivery of fiber optic sensing into a near-universal design, and enable the simple replacement of any damaged or deteriorated sensor-embedded pavement sections with new, identical sections. This may prevent the potential inexact on-site installations, especially when workers are working in harsh conditions and attempting to complete the work under time constraints. This concept is illustrated in Fig. 25. The networks of slabs may adopt the Plug and Play nature of the Internet.