In this study, a motion analysis system based on deep learning for sports training videos was developed to optimize the standardization of athletes’ movements and the scientific level of training. This scheme builds a hierarchical recognition framework by integrating 3D convolutional neural network and bidirectional timing modeling module: First, 3D-CNN is used to extract spatial-short-term motion features in video clips, then the gated loop unit is used to capture the long-range action association across frames, and finally, multi-head attention mechanism is used to strengthen the discriminant features of key action clips. In order to improve the robustness of the model, kinematic data enhancement strategies (horizontal flipping, timing clipping) and feature transfer methods based on Kinetics data sets were introduced in the training stage. The comparison experiments on UCF101 and HMDB51 standard test sets show that the system achieves 92.7% Top-1 accuracy in basketball shooting action recognition and other tasks, which is 19.3 percentage points higher than the traditional HOG + SVM method, and provides an effective algorithm basis for the construction of intelligent sports teaching assistance platform.