Applying Transfer Learning to Traffic Surveillance Videos for Accident Detection

Abstract

Automated traffic video surveillance is a crucial research domain in computer vision due to the need to enable highway safety. It is very important to detect road accidents from traffic surveillance videos in an automated manner to take necessary actions and save the lives of people and properties. Motivated by the same, this paper proposes a method to detect road accidents from traffic surveillance videos in an automated manner. Specifically, we use an object-centric accident detection model using the YOLOv2 architecture based on the transfer learning technique. The YOLOv2 model is a homogeneous convolutional architecture that makes it faster to predict bounding boxes. This work includes a brief description of the YOLOv2 architecture and how we fine-tune a 32-layer variant pre-trained on the VOC dataset to our custom accident dataset. Our experiments using a real-world anomaly detection dataset show significant results in terms of mean average precision. Moreover, our model works in real-time, achieving 60 FPS on an NVIDIA Tesla K80 GPU and ~16.67 FPS on a standard laptop with a 4GB GT GPU. Our implementation can thus provide a near real-time accident localization with 76% mAP on the road accident dataset.