||Children’s Football Action Recognition based on LSTM and a V-DBN
||(Zhaosheng Chen) ; (Na Chen)
|| Football; Action recognition; V-DBN; LSTM; Children’s features
||In order to improve teaching children how to play football, combining the Vector of Locally Aggregated Descriptors (VLAD) model and a deep belief network (DBN) into a V-DBN is proposed based on 3D bone recognition that recognizes football actions. We use the contrast method to reduce the dimensionality of action features, and we complete the action recognition through analysis of key parameters. After experimental testing with the MSRAction3D data set, Grassmann manifold and graph-based action classification and recognition reach accuracies of 79.2% and 93.4%, respectively, after 100 iterations of training, but the V-DBN reaches 98.6%. In the UTKinect-Action database test, the average recognition rates of Grassmann manifold and graph-based action classification and recognition are 88.38% and 91.31% accurate, respectively, while the VLAD is 93.96% accurate, showing the best overall performance. However, the effect in single-action recognition is only average. Using the LSTM optimization model on results from infant football action recognition, the average accuracy rate of LSTM+V-DBN is 0.981 compared to the V-DBN at 0.892. Clearly, the optimized LSTM+V-DBN model performs better in toddler action recognition. This research provides important reference value to the application of human action recognition technology in children’s football education.