YOLO-Based Object Detection and Tracking for Autonomous Vehicles Using Edge Devices

Abstract

One of the essential tasks for Autonomous Driving and Driving Assistance systems is the detection and tracking of Vulnerable Road Users (VRU) and traffic objects. Many recent developments in this area have been leveraging Deep Learning techniques. However, these Deep Learning models require heavy computational power. For this reason, optimising software components coupled with adequate hardware choices is crucial in the development of a system that can infer in real-time. This paper proposes solutions for object detection and tracking in an Autonomous Driving scenario by comparing and exploring the applicability of different State-of-the-art object detectors trained on the BDD100K dataset, namely YOLOv5, Scaled-YOLOv4 and YOLOR. In addition, the paper explores the deployment of these algorithms on Edge Devices, more specifically, the NVIDIA Jetson AGX Xavier. Furthermore, it examines the use of the DeepStream technology for real-time inference by comparing different object trackers, such as NvDCF and DeepSORT, in the KITTI tracking dataset. The proposed solution considers a YOLOR-CSP architecture with a DeepSORT tracker running at 33.3 FPS with a detection interval of one and 17 FPS with an interval of one.