Improved coordinate attention network for classification of dangerous driving behavior

Abstract

With the rise of traffic accidents caused by unsafe driving behaviors, the accurate classification of these behaviors has become a pressing issue in intelligent transportation systems. Traditional methods such as AlexNet and VGG, while effective for general image recognition tasks, fail to capture the complex and subtle features necessary for recognizing dangerous driving behaviors. To address this, we propose an improved residual network model, SC-ResNet, which integrates a coordinate attention mechanism and SIFT (Scale-Invariant Feature Transform) feature fusion to enhance classification accuracy under varying conditions including rotation, scale, and illumination changes. Furthermore, we introduce a multi-scale feature pyramid network and a novel joint loss function to better handle the multi-class classification imbalance problem. Experimental results show that our model outperforms traditional networks by $0.6\text{\%}$ to $4.7\text{\%}$ in classification accuracy. Future research will focus on improving model generalization and computational efficiency for real-time applications.