Distracted Driving Behavior Recognition System Based on Deep Learning Approach and Multi-View Imaging

Traditional single-view driver-monitoring systems often exhibit reduced detection efficiencies and accuracies owing to occlusions and viewpoint limitations. This challenge is particularly pronounced in multi-passenger driving environments, where a single camera may fail to capture critical driver behaviors, increasing the risk of distracted driving. Deep-learning models with high parameter counts face practical constraints related to hardware resources and computational demands in real-world applications. Consequently, developing lightweight and high-efficiency driver behavior recognition models has become a key research focus. This study proposes a lightweight temporal graph convolutional network based on multi-view feature fusion for distracted driver behavior recognition system. The model employs an adaptive multi-view attention module and a temporal-spatial attention module to effectively aggregate driver behavior cues from different viewpoints while filtering redundant temporal features, thereby enhancing recognition performance. Experimental results demonstrate that the proposed method achieves accuracies of 96.80% and 96.04% on the SynDD1 and DMD datasets, respectively, while reducing parameters and computational complexity to 1.10 M and 0.21 GFLOPs. These findings validate the effectiveness and practicality of the proposed approach, offering valuable insights for future research and applications in driver behavior recognition.