Real-Time Traffic Counting on Resource Constrained Embedded Systems

Real-time traffic monitoring is crucial to meet the demand for intelligent traffic coordination. Adjusted green phases can improve traffic flow and lead to congestion avoidance. Therefore, in addition to object counting, it is crucial for optimal traffic monitoring to track object movement. In this paper, we propose a highly optimized framework for detecting, tracking and counting road users at intersections. The nano YOLOv5 is used for detection and a neural architecture searched network is proposed for feature extraction of the DeepSORT algorithm. Both are compressed using TensorRT. Between detected frames, an optical flow-based tracker provides consistent processing of tracked objects and relieves the embedded system of some computationally intensive neural network inference. Motion and box shape features are taken into account to improve robustness in re-identification and prevent swapping boxes.