3D-CNN-based Action Recognition Algorithm for Basketball Players

Abstract

With the development of artificial intelligence, there are numerous analysis methods for human action recognition. In basketball, its technical action features are obvious, so the feasibility of recognizing and classifying its technical actions is high. However, due to the existing action recognition methods are difficult to effectively utilize continuous frames, resulting in poor accuracy of basketball technical action recognition. Thus, the study suggests a continuous frame action identification approach based on the single-shot multi-edge detection algorithm and 3D convolutional neural network in order to enhance the performance of technical action recognition. The experimental results revealed that single shot multibox detector algorithm accurately recognizes the human body in the image and labels its confidence level. In addition, in basketball action recognition, the loss value of original frame was 6.0 and 6.8 on the training set and validation set, respectively, and the loss value of crop frame was 5.1 and 5.9 on the training set and validation set, respectively. 3D convolutional neural network achieved the highest classification accuracy of about 88.3% for the stop-and-go jump shot action in the original frame and its crop frame with an average recognition rate of about 90.3%. The recognition accuracies of original frame and crop frame increased with the increase of epoch, and reached a stable state when the epoch was 30, due to the presence of variable features in the European step, change of direction, and Sam Gould's action, which led to misjudgment of both original frame and crop frame. The accuracies of the original frame training set and test set were 0.91 and 0.81, respectively, and the accuracies of the crop frame training set and test set were about 0.92 and 0.81, respectively. After the fusion of the original frame features and the crop frame features, the average recognition rate was about 94.6%, which was significantly higher than that of the single-resolution recognition. Recognition. In addition, with the increase of frame input, the F1-measure gradually increased, while the false positive rate gradually decreased. When the frame input was 7, the F1-measure and the misjudgment rate were 0.79 and 0.19, respectively. When the frame input was 16, the F1-measure and the misjudgment rate were 0.92 and 0.05, respectively. The above results show that the continuous frame action recognition method based on single-shot multi-frame detection algorithm and three-dimensional convolutional neural network can realize the accurate recognition of the technical action in basketball video.