Automatic Recognition of Driving Events based on Deep Learning

Abstract

In recent years, there has been a rapid development of intelligent driving assistance systems. Although most vehicles nowadays are equipped with driving assistance systems, the number of car accidents continues to rise. The main cause of car accidents is still largely attributed to human factors. Therefore, there has been an increasing focus on research related to accident detection and driver behavior analysis. This study used deep learning methods to automatically recognize driving events from recorded driving videos. In the training phase of deep learning, we cropped all the videos in the training data into multiple clips, and labeled driving event categories for each clip, including four categories: vehicle stopped, straight driving, turning, and collision. The proposed model references the architecture of the SlowFastNet model and the concepts of I3D. We expanded Inception-V3 to a 3D structure and replaced the bottom architecture of SlowFastNet with 3D-Inception-V3, making the network more applicable to the training data. After training, the model can recognize driving events in various driving environments. Through experimental comparisons, our network architecture achieved the highest recognition accuracy, with an accuracy rate of 93.3%.