Abstract: Aiming at the temporal characteristics of human behavior in videos, a video human beha- vior recognition method is proposed that combines asymmetric convolution and CBR modules. This method uses 3D ResNet-50 as the backbone network. First, the larger convolutions in the network are changed to the concatenation of two asymmetric 3D convolutions, which deepens the local key feature extraction of the convolution layer in the horizontal and vertical directions. Secondly, CBR module is added to improve the number of network layers. The network extracts multi-angle features of images and time series from continuous video frame sequences, classifies them according to the feature data, and finally outputs the recognition results. Extensive experimental results on the benchmark dataset UCF101 show that the Top1 and Top5 accuracy of the proposed method are improved by 4.03% and 4.99%, respectively, compared with the original 3D ResNet network, and the recognition accuracy of this method is also better than other mainstream methods.

Key words: human behavior recognition, 3D convolution, 3D ResNet, asymmetric convolution, UCF101 dataset