关键词: 智能交通, 计算机视觉, 事故检测, 域适应双流网络

Abstract: Road traffic accidents are one of the major causes of property damage, casualties, and traffic congestion. Rapid and accurate detection of accidents is crucial for improving emergency response speed and reducing the fatality rate. In recent years, vision?based accident detection methods have become the mainstream, achieving significant progress through the application of deep learning techniques. However, existing methods suffer from low accuracy in domestic surveillance scenes due to significant domain shift between the open-source training data and domestic surveillance accident video data. To avoid the laborious task of constructing and annotating a large-scale domestic surveillance accident dataset, this paper proposes a domain adaptation dual-stream network that fully utilizes existing open-source accident datasets to achieve road accident detection in domestic surveillance scenes. Specifically, this paper develops a coarse-to-fine progressive domain adaptation training approach to guide the model in capturing domain-invariant accident appearance features. After that, a cross-attention fusion network is employed to establish correlations and interactions between appearance and motion features, alleviating feature discrepancies caused by scene variations. Experimental results demonstrate that the proposed method achieves high detection performance in domestic surveillance scenes, validating its effectiveness and reliability. This research is expected to provide important technical support for enhancing public safety and improving traffic efficiency.

Key words: intelligent transportation, computer vision, accident detection, domain adaptation dual-stream network