Deep Learning Envisioned Accident Detection System

Abstract

The conflict between computational overhead and detection accuracy affects nearly every Automated Accident Detection (AAD) system. Although the accuracy of detection and classification approaches has recently improved dramatically, the systems’ high computational resource requirements make them unsuitable for deployment in situations when immediate feedback is required. This paper suggests a strategy for developing an accurate, cost-efficient automatic collision identification system that can be deployed with a minimal amount of hardware. The AAD system is divided into three key phases—Detection, Tracking, and Classification and offers computationally reduced intensive algorithms for each of these stages. For the detection phase, YOLOv3, it uses a deep learning model that has been trained through knowledge distillation. Its accuracy is on par with that of YOLO (You-Only-Look-Once). On the basis of the information inference and data mining techniques COCO (Common Objects in Context) dataset, YOLOv3 surpasses all other identification techniques considering runtime complexity, averaging an astounding 30 frames/second on a low-end system. It attains an Average Precision (AP) mark of 44.8. For the tracking phase, the system makes use of SORT (Simple Online Real-time Tracking). Since a radial basis kernel with a support vector machine works efficiently, an area under the curve (or AUC) score of 0.92 is achieved. Alongside this, this paper examines a number of machine-learning techniques for the classification stage.