Long-Range Multi-Object Tracking at Traffic Intersections on Low-Power Devices

The next generation of intelligent traffic signal control systems needs multi-object tracking (MOT) algorithms that can track vehicles hundreds of meters away from traffic intersections. To facilitate the integration of long-range MOT into existing traffic infrastructure, the tracker must achieve a good balance of cost-effectiveness, accuracy, and efficiency. Although much progress has been made on deep-learning-based MOT for video, these approaches have limited applicability for edge deployment since deep neural networks typically require power-hungry hardware accelerators to achieve real-time performance. Furthermore, traffic cameras have a field of view limited to near the intersection. To address these shortcomings, we introduce a practical MOT framework that fuses tracks from a novel video MOT neural architecture designed for low-power edge devices with tracks from a commercially available traffic radar. The proposed neural architecture achieves high efficiency by using depthwise separable convolutions to jointly predict object detections alongside a dense grid of features at a single scale for spatiotemporal object re-identification. A simple and effective late fusion strategy is also presented where tracks of distant vehicles from a traffic radar are handed over to the video tracker within a region where the sensor fields of view overlap. Our video tracker is empirically validated on the UA-DETRAC video MOT benchmark for traffic intersections and the multi-sensor tracker is evaluated on video and radar data collected and labeled by the authors at an instrumented traffic intersection.