3D Convolutional Enhanced Driver Human and Object Interaction Behavior Recognition
The behavior of drivers smoking and making phone calls is a typical human interaction task,in order to improve the model's resistance to occlusion and light changes in the driving environment,as well as the accuracy of human interaction behavior matching.This study first proposes a 2D expanded group attention mechanism method to optimize the target detection network and improve the small target detection performance of human and object paths;Then,a high-precision lightweight module is proposed that combines the 3D expanded group attention mechanism with 3D group convolution to construct a dynamic video behavior recognition model and enhance the nonlinear feature extraction ability of temporal space;Finally,the inter frame intersection and union ratio statistical judgment results of the image are combined with the predicted results of the dynamic video behavior recognition model to make the final driver character interaction behavior judgment.The experimental results demonstrate the effectiveness of 2D and 3D expanded group attention mechanisms in behavior recognition,with an average accuracy and recall improvement of 12.5％and 7.72％in driver character interaction behavior.Especially in scenarios where cigarettes and mobile phones are obstructed or the lighting conditions are unfavorable,the improvement is significant,and it can solve the problem of confusion and recognition between the driver and their rear passengers'behavior.
NETL
NSTL
万方数据
维普
