Real-time image enhancement for an automatic automobile accident detection through CCTV using deep learning

Abstract

Almost all of the automatic accident detection (AAD) system suffers from the tradeoff between computational overhead and detection accuracy. Recent advances in detection and classification methodologies have shown phenomenal improvements in accuracy but these systems require a huge number of computational resources making them unviable for deployment requiring real-time feedback. This paper proposes a methodology to develop a reliable and computationally inexpensive real-time automatic accident detection system that can be deployed with minimum hardware requirements. Specifically, we split our AAD system into three major stages (Detection, Tracking and Classification) and propose algorithms for each stage with reduced computational need. For the detection stage, we propose Mini-YOLO, a deep learning model architecture trained using knowledge distillation that has comparable accuracy with its counterpart YOLO(You-Only-Look-Once) with reduced model size and computational overhead. Mini-YOLO achieves an average precision (AP) score of 34.2 on the MS-COCO dataset while outperforming all other detection algorithms in runtime complexity, achieving a staggering 28 frames per second on a low-end machine. For the tracking stage, we adopt SORT (Simple Online Real-time Tracking) and for classification stage, we compare multiple machine learning algorithms and show that a support vector machine with radial basis kernel performs the best with an area under the curve (AUC) score of 0.98, model size of 448 KB (kilobytes) and 12.73 ms (milliseconds) latency.