Vehicle detection and tracking systems collect continuously generated traffic information, such as vehicle counting, speed measurement, and traffic analysis. These processes can be time-consuming, and traditional human-based systems have a relatively high percentage of errors due to a lack of concentration and experience. Researchers Ahmad H. Alomari and Enas Abu Lebdeh hypothesized that a computer-based system using computer vision and machine learning algorithms would offer greater accuracy and efficiency.
Their paper, “Smart Real-Time Vehicle Detection and Tracking System Using Road Surveillance Cameras” in the Journal of Transportation Engineering, Part A: Systems, proposes using different computer vision algorithms for vehicle detection and tracking. The authors introduced a complete system to handle different applications, including detection, classification, recognition, tracking, and related statistic registration. Learn more about how this research could improve road usage and traffic analysis at https://doi.org/10.1061/JTEPBS.0000728. The abstract is below.
Abstract
Several traffic studies necessitate vehicle counting during peak hours and throughout the day, as well as detailed classification and tracking, which consumes human time and effort, particularly at intersections. Manual efforts primarily collect the necessary traffic demand data live in the field or from video recordings using an extensive data manipulation process. Alternative solutions include computer-based systems that perform human tasks more efficiently and with less time and effort, and these systems vary in function and performance. This paper proposes a comprehensive computer-based system that detects, tracks, and computes related statistics in real time while making the best use of available resources, such as public road surveillance cameras. The main contribution of this work is the effectiveness of combining various computer vision algorithms to achieve high-accuracy performance during real-time streaming of road cameras. The experiments confirm the system’s performance by achieving an average success rate of 93.2%. The novel aspect of this work is that detections, point extractions, matching, tracking, and classification were implemented in a single system that ensures real-time execution and high-accuracy output and uses existing infrastructure. The system compensates for the variations in light between day and night and between cloudy and sunny weather. It also recovers hidden vehicles and changes in view for each vehicle as it moves. The proposed method efficiently and partially integrates some mechanisms into a single system.
Read the paper in full in the ASCE Library: https://doi.org/10.1061/JTEPBS.0000728.
