PaperlayerView exact revision on arXiv

Source-linked AI summary

Intelligent Traffic Monitoring Systems for Vehicle Classification: A Survey

Myounggyu Won

arXiv:1910.04656v2cs.CYeess.SP

TL;DR

Traffic agencies need vehicle-classification technologies to support traffic analysis, roadway planning, and safety, but emerging systems differ substantially in sensors, configurations, environments, and costs. This paper surveys and organizes these systems, analyzes their designs and challenges, and identifies open problems and future directions. Its synthesis covers broad classes of sensing approaches and is intended to support technology selection by academia, industry, and government agencies.

  • Problem

    Emerging vehicle-classification systems differ in sensors, hardware settings, configurations, operating environments, and costs, making appropriate technology selection difficult.

  • Method

    The paper surveys state-of-the-art vehicle-classification technologies, categorizes them by deployment location and sensor or analysis approach, and reviews their designs, deployment issues, and challenges.

  • Results

    The survey synthesizes in-roadway, over-roadway, and side-roadway systems, including magnetic-sensor approaches, camera and infrared systems, reported performance, and future research challenges.

  • Takeaways & Limitations

    The review is intended as a resource for selecting appropriate vehicle-classification solutions across academic, industrial, and governmental traffic-monitoring applications.

Abstract

from arXiv · show

A traffic monitoring system is an integral part of Intelligent Transportation Systems (ITS). It is one of the critical transportation infrastructures that transportation agencies invest a huge amount of money to collect and analyze the traffic data to better utilize the roadway systems, improve the safety of transportation, and establish future transportation plans. With recent advances in MEMS, machine learning, and wireless communication technologies, numerous innovative traffic monitoring systems have been developed. In this article, we present a review of state-of-the-art traffic monitoring systems focusing on the major functionality--vehicle classification. We organize various vehicle classification systems, examine research issues and technical challenges, and discuss hardware/software design, deployment experience, and system performance of vehicle classification systems. Finally, we discuss a number of critical open problems and future research directions in an aim to provide valuable resources to academia, industry, and government agencies for selecting appropriate technologies for their traffic monitoring applications.

I. INTRODUCTION

Traffic monitoring systems support congestion mitigation, roadway utilization, transportation planning, and safety, but the growing range of vehicle-classification technologies makes selecting an appropriate solution difficult. This survey organizes these systems by deployment location, reviews their technologies and challenges, and identifies open research directions.

  • Motivation: Traffic monitoring systems collect vehicle counts, types, and speeds to support roadway utilization, transportation planning, and transportation safety.Transportation agencies invest substantially in developing, deploying, and maintaining these systems.
  • Motivation: Vehicle classification is crucial because vehicle-type information supports highway capacity estimation and pavement maintenance planning.
  • Research need: Recent sensing, machine-learning, and wireless-communication advances have increased classification accuracy while also diversifying sensors, configurations, operating environments, and costs.This diversity complicates technology selection for transportation applications.
  • Survey scope: The survey reviews state-of-the-art classification technologies, organizes them by sensor type and analysis mechanism, compares their research and technical issues, and offers selection guidelines.It also discusses hardware/software design, deployment experience, open problems, and future directions.
  • Taxonomy: Vehicle classification systems are categorized by deployment location as in-roadway-based, over-roadway-based, or side-roadway-based systems.The taxonomy is further refined using sensor types and methods for analyzing and utilizing sensor data.
  • Taxonomy: Over-roadway systems can cover multiple lanes simultaneously, whereas camera-based systems are affected by weather, lighting, and privacy concerns.Infrared and laser-scanner systems are discussed as alternatives addressing privacy concerns.

III. IN-ROADWAY-BASED VEHICLE CLASSIFICATION

In-roadway systems classify vehicles using sensors installed on or beneath pavement, extracting features such as axle configuration, magnetic signatures, vehicle length, and seismic signals. The section reviews loop detectors and other sensor systems, emphasizing accuracy, cost, and technical limitations.

  • III. IN-ROADWAY-BASED VEHICLE CLASSIFICATION: In-roadway classification systems use pavement-installed sensors to extract axle, length, waveform, and other vehicle features.The reviewed systems include loop, piezoelectric, magnetometer, vibration, and related sensor configurations.
  • A. LOOP DETECTORS: Loop detectors generate vehicle-dependent magnetic profiles from changes in inductance, with signal amplitude, phase, and frequency supporting classification.Loop detectors may be saw-cut or preformed and may use single or dual loops.
  • A. LOOP DETECTORS: Vehicle classes with similar axle configurations or body lengths remain difficult to distinguish accurately.The reviewed systems address this using body signatures, multi-parameter classifiers, or additional sensing features.
  • A. LOOP DETECTORS: Single-loop classification can reduce cost, while dual-loop systems provide speed and length measurements but are more expensive.Single-loop approaches reported about 96% accuracy for cars, trucks, and vans, and 99.4% for long-vehicle detection in separate studies.
  • A. LOOP DETECTORS: Over 98% average accuracy was achieved when dual-loop systems classified vehicles into three length classes with boundaries at 28 ft and 46 ft.The method accounts for non-zero acceleration when estimating vehicle length.

B. MAGNETIC SENSORS

Magnetic sensor systems classify vehicles from magnetic-field disturbances caused by ferrous vehicle components. The reviewed approaches use sensor networks, single-sensor waveform learning, or hybrid measurements combining magnetic signatures with vehicle length or other features.

  • B. MAGNETIC SENSORS: Magnetic sensors detect vehicle-dependent disturbances in the Earth’s magnetic field and offer advantages in size, weight, cost, and energy efficiency.These systems classify vehicles from distinctive magnetic-field changes produced by different vehicle bodies.
  • B. MAGNETIC SENSORS: Magnetic classification systems comprise multi-sensor networks based on vehicle length, single-sensor waveform analysis, and hybrid systems combining both information sources.This taxonomy organizes the principal architectures reviewed in the section.
  • B. MAGNETIC SENSORS: Over 97% accuracy was reported for passenger vehicles, single-unit trucks, combination trucks, and multi-trailer trucks using vehicle magnetic length.The feature is derived from vehicle speed and the time spent over a magnetic sensor.
  • B. MAGNETIC SENSORS: 96.4% average accuracy was achieved by fusing magnetic waveforms from two sensors 80 m apart to classify four vehicle types.The approach uses waveform segmentation and sensor fusion to improve classification.
  • B. MAGNETIC SENSORS: Single-sensor models use automatically extracted waveform features, including speed-independent features and statistical, energy, and short-term features.Reported approaches include CART and XGBoost classifiers for multiple vehicle categories.

C. VIBRATION SENSORS

Vibration sensors capture localized pavement vibrations generated by passing vehicles, but classification is complicated by geology and complex seismic waveforms. Reviewed systems use axle features, seismic-signal features, and machine learning to classify vehicles.

  • C. VIBRATION SENSORS: Vibration sensors use pavement as a transducer to capture localized vibration patterns induced by passing vehicles.The signals vary with the propagation effects of the underlying geology and with waveform form, direction, and speed.
  • C. VIBRATION SENSORS: Vibration-based systems mainly classify vehicles from axle count and spacing or from features extracted from induced seismic waveforms.Machine learning is often used because seismic signals are complex.
  • C. VIBRATION SENSORS: A hybrid system uses magnetic sensors for vehicle arrival and departure times, while vibration sensors estimate axle count and axle spacing.These measurements form the key classification features.
  • C. VIBRATION SENSORS: 89.4% average accuracy was achieved by DOVS across 10 vehicle types.The system adds frequency-domain features and vehicle speed to distinguish vehicles with similar axle configurations.
  • C. VIBRATION SENSORS: 92% accuracy was the best result reported for classifying assault amphibian vehicles and dragon wagons from seismic signals.The system used vibration data and machine learning on two vehicle classes.

D. OTHER TECHNOLOGIES

Other in-roadway technologies classify vehicles using roadway-embedded sensors and features such as axle count, axle spacing, vehicle length, and weight. These systems include integrated, piezoelectric, and fiber-optic sensor approaches.

  • In-roadway systems use weigh-in-motion, piezoelectric, and fiber-optic sensors for vehicle classification.
  • Combining vehicle weight with axle spacing supports classification using multiple classifier models.The evaluated models include Naive Bayes, Decision Tree, SVM, and neural-network classifiers.
  • A 16-element piezoelectric system uses tire count, vehicle length, and axle spacing as classification features.Vehicle speed is estimated from impact timing across sensors aligned with the lane axis.
  • Two fiber Bragg grating sensors extract axle count and axle spacing from pavement strain signals.The sensors capture strains generated when vehicles pass over the roadway.

IV. OVER-ROADWAY-BASED VEHICLE CLASSIFICATION

Over-roadway systems provide non-intrusive, multi-lane vehicle classification, with cameras predominating because they capture rich visual and geometric information. Research addresses image capture, feature extraction, classification, occlusion, data variability, and privacy-related alternatives.

  • Over-roadway systems avoid physical roadway changes, reducing construction and maintenance costs while potentially covering multiple lanes or entire road segments.
  • A. CAMERAS: A single camera can classify vehicles across multiple lanes using visual features and vehicle geometry.
  • A. CAMERAS: Camera-based classification captures vehicle images, extracts features, and applies an algorithm, with methods differing in image capture, feature type, and classification mechanism.
  • A. CAMERAS: 97.7% classification accuracy was achieved on the 11-class MIO-TCD dataset using augmentation and an ensemble of CNN models.
  • A. CAMERAS: 99.0% accuracy was reported for a hybrid camera–optical-sensor classifier combining CNN and optical-sensor decisions with Gradient Boosting.

B. AERIAL PLATFORMS

Aerial platforms such as UAVs and satellites provide wide-area coverage for vehicle classification, but low image resolution makes detection and fine-grained classification difficult. Reported systems therefore often use limited vehicle classes and achieve lower accuracy than ground-based systems.

  • Aerial platforms can cover wide roadway areas, but low image resolution makes even vehicle detection difficult.One airborne system achieved 90% detection accuracy, compared with near 99% detection accuracy for most ground-based systems.
  • Many aerial systems classify only easily distinguishable categories such as cars and trucks because of low image resolution.
  • 80.3% average classification accuracy was achieved for sedans, vans, pickups, and trucks using change detection followed by CNN classification.
  • VGG-16 achieved the highest classification accuracy among the reported LeNet, AlexNet, and VGG-16 aerial-image models.The cited passage reports the highest accuracy as 80%.

C. PRIVACY PRESERVING SOLUTIONS

Privacy-preserving vehicle classification replaces cameras with infrared, ultrasonic, or laser-scanning sensors. These approaches avoid camera privacy concerns but introduce performance, cost, or environmental trade-offs.

  • Combining infrared and ultrasonic sensors classified sedans, pickup trucks, SUVs, buses, and two-wheelers with up to 99% accuracy.The system fused sensor features using Bayesian and neural networks.
  • Laser scanners preserve privacy by estimating vehicle width, height, and length from three-dimensional surface scans.A laser-scanner profile was represented as an image for classification, reaching 82.5% accuracy across six vehicle types.
  • Laser scanners are sensitive to extreme weather and cost more to install than cameras.
  • Vehicle-shape domain knowledge and alignment methods were used to improve laser-scanner classification.

V. SIDE-ROADWAY-BASED VEHICLE CLASSIFICATION

Side-roadway systems place sensors beside the road, enabling simultaneous coverage of multiple lanes without traffic disturbance or lane closures. Their main challenge is classifying overlapping or occluded vehicles.

  • Side-roadway systems can monitor multiple lanes simultaneously while avoiding traffic disturbance and lane closure.This makes them especially appropriate for ad-hoc monitoring purposes.
  • Occlusion distorts sensor data and makes detecting and classifying overlapping vehicles especially challenging.Data from front vehicles can also be distorted by vehicles behind them.
  • Side-roadway implementations use magnetic, acoustic, LIDAR, radar, radio, and Wi-Fi transceiver sensors.

A. MAGNETIC SENSORS

Magnetic sensors support side-roadway vehicle classification by providing vehicle dimensions and signal features, while collaborative sensing and specialized models address difficult traffic conditions and deployment constraints.

  • Side-roadway magnetic sensors help reduce the installation and maintenance costs associated with in-road systems.
  • Three-axis magnetic sensing uses vehicle length and height to classify vehicles with similar body sizes.Unlike systems measuring only vehicle length, this approach also obtains vehicle height information.
  • A hierarchical tree-based approach extracts five magnetic-signal features for classifying vehicles traveling close together in low-speed congested traffic.The features include signal duration, signal energy, average energy, and positive-to-negative energy ratios.
  • Collaborative magnetic-camera sensing activates the camera only after vehicle detection to reduce power consumption.
  • Magnetic-sensor systems use varied features and models, including duration, vehicle-body information, RSSI, kNN, and SVM.

B. ACOUSTIC SENSORS

Acoustic systems classify vehicles from audio signals but are sensitive to ambient noise, so they are often combined with other sensors or deployed as sensor networks.

  • Acoustic classification depends on effective feature extraction and is strongly affected by ambient noise.Consequently, acoustic sensors typically support other sensing modalities.
  • An audio classifier can supervise autonomous training of a video classifier, removing the need to manually label large amounts of video data.The audio system forwards a priori classifications and confidence levels to the video system.
  • A wireless acoustic sensor network combines sensor-level decisions and detects faulty sensors using multiple microphones.
  • 96.3% average classification accuracy was reported on the DARPA/IXOs SensIT dataset for Assault Amphibian Vehicle and Dragon Wagon classes.

C. LIDAR

LIDAR and related roadside sensing systems classify vehicles from body geometry or signal reflections, while RF and Wi-Fi approaches offer weather resilience or lower-cost deployment alternatives.

  • LIDAR extracts vehicle size and shape features with high-precision sensing, although vehicle occlusion remains a challenge.
  • A roadside system uses two LIDAR sensors to vertically scan vehicle bodies and extract geometric features for classification.The features include vehicle height, length, middle-drop measurements, and front and rear body dimensions.
  • Truck body types can be classified from vehicle-body-point duration and shape using DT, ANN, SVM, and NB classifiers.The system was deployed roadside to classify five truck body types.
  • Radar uses radio-wave reflections and is less vulnerable to weather and light than LIDAR, while LIDAR provides a more accurate vehicle-body representation.
  • RF classification uses signal attenuation or RSSI patterns caused by a passing vehicle, with one system reporting 99% accuracy for passenger cars and trucks.
  • Wi-Fi systems exploit CSI patterns for low-cost deployment, achieving 96% average accuracy in an initial prototype and later supporting more vehicle types with CNNs.

VI. CHALLENGES FOR FUTURE RESEARCH

The review identifies unresolved challenges in fairly evaluating and improving vehicle classification systems as technologies and deployment conditions vary. It highlights standardization, occlusion, data requirements, heterogeneous sensing, and emerging V2X integration as directions for future research.

  • Evaluation standards: Vehicle classification systems need standardized vehicle-type lists so their performance can be evaluated more fairly and users can compare alternatives.The review notes that systems evaluated with fewer vehicle types tend to report higher accuracy, which is not guaranteed with more types.
  • Evaluation standards: Fair comparison also requires standardized experimental conditions covering lanes, obstacles, weather, and sensor-specific environmental effects.Weather, lane count, and ambient noise can affect different sensing technologies in different ways.
  • Performance metrics: Evaluation should include installation and maintenance cost, overlapped-vehicle capability, operational sustainability, weather and noise resilience, and privacy rather than accuracy alone.Camera-based systems may raise privacy concerns, while in-road systems can be costly to install and maintain.
  • Deployment challenges: Side-roadway systems face vehicle occlusion because vehicles can interrupt magnetic, LIDAR, radar, RF, and Wi-Fi sensing across lanes.The review proposes placing side-firing sensors at different heights to target lanes explicitly.
  • Learning and adaptation: Machine-learning systems require substantial training data, manual labels, and ground truth, motivating closed-loop self-learning models that evolve after deployment.The proposed direction is autonomous and continuous model evolution through trial and error.
  • Integrated and connected systems: Near-100% accuracy remains difficult with many vehicle types, while hybrid sensor and deployment systems may combine complementary strengths; V2X systems also require reliable, secure communication protocols.The review reports no identified optimal integration method for heterogeneous classification systems and lists protocol challenges for mixed V2X and traditional traffic.
Loading 1910.04656v2…