Combining the SSD Target Identification Algorithm with the 3d-Cnn Architecture for Transfer Learning Research in Basketball Training

Abstract

The development of deep learning and artificial intelligence has made the large amount of data generated by various types of human actions of great analytical value. The continuous updating of recognition algorithms based on text and picture frames has also made the movement recognition in video of some research value. Currently, there are few studies on technical movement recognition in basketball. Based on this, this study tests the performance of the constructed target detection algorithm and movement recognition algorithm. The experimental results found that the maximum detection accuracy of Fast R-CNN, YOLO, and SSD algorithms on basketball dataset were 85.9, 84.9, and 93.8, respectively. in addition, the recognition accuracy of ( 3D Convolutional Neural Network, 3D-CNN ) 3D-CNN and dual-resolution 3D-CNN were compared under different video frames. When the quantity of video frames is 20, the two algorithm models have the highest recognition accuracy of basketball basic movements, 89.6 and 95.8, respectively.