Improved C3D Network Model Construction and Its Posture Recognition Study in Swimming Sports

Abstract

In order to solve the problem of low recognition accuracy caused by the C3D network limited by the large number of parameters, the study proposes an improved C3D network-based pose recognition model. The improvement of the C3D network is realized by using global average pooling instead of fully connected layer, and the attention residual network on the basis of improved C3D is further designed, and the attention staged residual network model is constructed by introducing the spatio-temporal channel attention mechanism. Comparative validation showed that the improved C3D network increased the accuracy by 13.49% over the C3D network on the HMDB51 dataset. When the various models were compared, it was found that the suggested model, which had an area under the receiver operating characteristic curve as high as 0.98, improved the study's accuracy over the two well-known networks by an average of 14.34%. The accuracy of the proposed model increased the accuracy of the study over the popular networks by an average of 14.50% for the recognition of the postures of all the swimming categories in the homemade swimming sports dataset. The findings show that the number of parameters in the enhanced C3D network proposed in the study has been successfully reduced, and that the attention residual network model based on the enhanced C3D network has a superior application value in sports pose recognition. It also offers some advantages in terms of fine-grainedness and recognition accuracy.