Sensing and Imaging | 2021
A Lightweight Two-stream Fusion Deep Neural Network Based on ResNet Model for Sports Motion Image Recognition
Abstract
Sports motion recognition aims to track the movement of some key points in the time domain and record human movement, which is of great significance for competitive training and national fitness. However, the traditional motion recognition methods have the disadvantages such as low recognition rate, time-consuming due to the vast model parameters and edge artifacts, etc. By combining the shallow multi-scale network with the deep network, this paper proposes a novel sports motion image recognition based on a lightweight two-stream fusion deep neural network with the ResNet model. This model can greatly reduce the number of model parameters and improve the accuracy of feature extraction. Finally, we conduct the experiments on four public data sets: UCF101, HMDB51, MSR3D and UCF-Sport. Experiment results show that the proposed algorithm has higher recognition accuracy and stability with various complex actions compared with other state-of-the-art motion recognition methods.