Accurate analysis of table tennis posture based on 3D convolutional neural network

Traditional pose resolution methods face the problem of insufficient accuracy when dealing with complex table tennis sports scenes. For this reason, this paper firstly extracts the foreground of table tennis movement based on the improved AlphaPose algorithm. Then a three-dimensional stacked convolutional neural network (3D-SCNN) is designed to extract the spatio-temporal features of the motion pose. A 3D convolutional block attention mechanism is proposed to enhance the important spatio-temporal features and output the final motion pose estimation results through fully connected layers. Finally, a similarity calculation method is used to calculate the cumulative distances of the key points in the standard and estimated movements, and scoring is performed, so as to realize the accurate analysis of the movement postures. The experimental outcome implies that the correct attitude estimation rate of 3D-SCNN is improved by at least 2.6%, which validates the effectiveness of the designed approach.