To construct a video action recognition model with 2D neural network speed while maintaining the performance of 3D neural network

the 3D multi-branch aggregation lightweight network action recognition algorithm is proposed. Firstly

the neural network is divided into multiple branches by using grouped convolution. Secondly

to promote the information exchange between branches

a multiplexer module with information aggregation function is added. Finally

the adaptive attention mechanism is introduced to redirect channel and spatio-temporal information. Experiments show that

the computational cost of the algorithm on the UCF101 dataset is 11.5GFlops

and the accuracy is 96.2%; the computational cost on the HMDB51 dataset is 11.5GFlops

and the accuracy is 74.7%. Compared with other action recognition algorithms

it improves the efficiency of the video recognition network and reflects certain recognition speed and accuracy advantages.