A method based on TSM for ego-crash recognition in dashcam videos

The scale of traffic accidents is growing, and they can cause massive casualties and property damage. Video data recorded in dashcams is the most important carrier for recording incident moments. Analyzing this video data holds significant importance. Advances in deep convolutional neural networks (CNNs) have significantly propelled the progress of visual accident recognition in the past. Compared to traditional 2D CNNs, 3D CNNs can effectively capture spatial temporal features. However, 3D CNNs suffer from computational issues. To improve the accuracy of 2D CNNs for accident recognition, we incorporate TSMs into 2D CNNs so they can capture more effectively the surface and motion features of traffic accidents in dashcam video data. This enables simultaneous learning of spatial-temporal features in 2D CNNs. We also incorporate coordinate attention into the model, enhancing its capability for spatial-temporal feature learning and improving model performance. Finally, our model achieves higher accuracy in accident recognition than 3D CNNs on a re-organized public traffic video dataset.