Cross-Modality Interaction-Based Traffic Accident Classification

Abstract

Traffic accidents on the road lead to serious personal and material damage. Furthermore, preventing secondary accidents caused by traffic accidents is crucial. As various technologies for detecting traffic accidents in videos using deep learning are being researched, this paper proposes a method to classify accident videos based on a video highlight detection network. To utilize video highlight detection for traffic accident classification, we generate information using the existing traffic accident videos. Moreover, we introduce the Car Crash Highlights Dataset (CCHD). This dataset contains a variety of weather conditions, such as snow, rain, and clear skies, as well as multiple types of traffic accidents. We compare and analyze the performance of various video highlight detection networks in traffic accident detection, thereby presenting an efficient video feature extraction method according to the accident and the optimal video highlight detection network. For the first time, we have applied video highlight detection networks to the task of traffic accident classification. In the task, the most superior video highlight detection network achieves a classification performance of up to 79.26% when using video, audio, and text as inputs, compared to using video and text alone. Moreover, we elaborated the analysis of our approach in the aspects of cross-modality interaction, self-attention and cross-attention, feature extraction, and negative loss.