A hybrid vehicle tracking System for Low-power Embedded Devices

The management of transportation networks becomes a major challenge in our cities. Different solutions are proposed to provide reliable traffic management services to improve road safety and traffic control. Traffic surveillance systems collect and analyze live footage from cameras to provide valuable information for monitoring. Vehicles tracking is one of the crucial tasks in these systems, as it allows the identification, tracking, and also the counting of traveling vehicles. Our objective in this work is to propose an optimized vehicle tracking algorithm that can operate seamlessly on low-cost embedded devices. This algorithm is intended to be the core processing part of a distributed framework of traffic surveillance. Our tracking algorithm is based on an adaptive hybrid tracking-by-detecting approach that combines both motion-based detection by the means of Background subtraction, and CNN-based detection using YOLO. The goal of this combination is to make the processing faster enough to be able to run on a Raspberry Pi 4. The idea behind our proposition is to eliminate the utilization of unnecessary computationally complex CNN-based detections. The experimental results demonstrate the effectiveness of our algorithm on Raspberry Pi, by achieving a matching rate above $90 \%$, with an average speed of about 16 FPS.