Deep Learning-based Visual Risk Warning System for Autonomous Driving

Abstract

In autonomous driving, the identification and tracking of multiple vehicles on the road are critical tasks. This paper aims to develop a risk warning system using deep learning algorithms to address the heterogeneous, high-dynamic, and complex driving environments. To enhance the generalization capability and detection accuracy of small objects in road perception, we propose a novel VBFNet-YOLOv8 algorithm for real-time vehicle identification, tracking, distance measurement, and speed estimation. Specifically, we replace the Backbone of the original YOLOv8 network with the VanillaNet structure and upgrade the traditional PANet in the neck part to Bi-FPN. By integrating the optimized YOLOv8n algorithm with Deepsort and TTC algorithms, we achieve a comprehensive road risk assessment. The algorithm continuously tracks the targets, and the TTC algorithm intuitively assesses the risk. Finally, the system provides layered warnings by changing the color of the bounding boxes, offering drivers an integrated and real-time risk alert. Comparative experimental results show that the optimized algorithm improves Precision by 0.61%, mAP@0.5 by 0.63%, and mAP@0.5:0.95 by 0.70%. In the road tests on sections A and B, the detection frame rate of the risk warning system maintained a minimum of 37.1fps and a maximum of 56.4fps. The detection Confidence of various objects remained above 0.67, reaching up to 0.97.