A data-driven approach for road accident detection in surveillance videos

The use of machine learning and computer vision techniques for detecting road accidents is a challenging task due to the limited availability of accident data for training. Staging fake accidents with real cars is expensive, and car crashes are rare incidents in roadside CCTV footage. Therefore, simulating fake car crashes using computers can be a feasible option. As such, we look at the following question in this paper; how successful can manually generated fake accident data be in terms of enabling a machine learning algorithm to detect real accidents?. In this work, we manually construct fake accident video frames from normal video traffic footage by creating simulated accidents. We do so by following predefined principles that maintain consistency with the scene context of normal frames. In order to detect real accidents in video footage, we fine-tune pre-trained deep convolutional neural networks on the manually generated fake accident frames. We use four pre-trained models i.e., AlexNet, GoogleNet, SqueezeNet and ResNet-50 on both normal and abnormal traffic video frames during the learning phase. The experimental results show that the fine-tuned AlexNet outperforms other models providing an 80% percent true positive rate when detecting anomalies (accidents) in real-world surveillance videos of UCF-Crime dataset. This demonstrates the validity of our hypothesis that simulated accident data could be valuable for training machine learning algorithms to detect real-world accidents.