Dual-stream architecture and improved action recognition based on 3D convolutional neural network fusion

Due to the rich content of video data and the excessive parameters of 3D convolutional networks, how to reduce the difficulty of parameter tuning and reduce the training time is a difficult problem in action recognition. To solve these difficulties, a new action recognition algorithm based on dual-stream architecture and improved three-dimensional convolutional aggregation is proposed. The network structure of the 3D convolution kernel is split into two types of convolution kernels, spatial flow and time flow, and the data flow is fused based on the dual-stream network, and the residual network is introduced for optimization. To verify the effectiveness of the algorithm, it is verified on the large public behavior dataset UCF101. Experimental results show that the algorithm has good spatiotemporal feature modeling ability, which effectively improves the accuracy of action recognition while shortening the training time.