The Bigger Picture: Scene Understanding for Recognizing Accidents within Mixed-Traffic Scenario using Deep Learning

Detecting road traffic accidents in a timely manner is crucial for reducing traffic fatalities. Computer vision techniques are increasingly being used for road traffic accident recognition because of their ability to automatically analyze traffic scenes without needing manual human interpretation. To date, Deep learning has made significant paces in addressing time-series problems and is increasingly being utilized for road traffic accident detection. However, detecting accidents in a mixed traffic flow environment is challenging due to difficulties in detecting smaller-sized road users and annotating unsafe trajectories, particularly in scenarios where larger-sized road users overshadow smaller-sized road users. Therefore, it may be more effective to consider the entire traffic scene as a whole, instead of focusing on individual objects. This paper proposes a Deep Learning model that analyzes the entire traffic scene using a recurrent neural network, ConvLSTM2D, for road traffic accident detection in mixed traffic flow environments. The model is evaluated on 139 videos of a testing dataset, achieving an accuracy of 84%. Future work will focus on further analyzing potential accident features across video frames to enhance the model’s performance.