<p>Traditional police training motion recognition often relies on single-modal visual data, which is easily affected by environmental conditions. There is an urgent need to develop a high-accuracy motion trajectory recognition model to support standardized evaluation and error correction during training. This can improve both training efficiency and field performance. This paper proposes a motion trajectory recognition model for police training that integrates multimodal perception and spatiotemporal features. The model collects Red-Green-Blue images, optical flow, grayscale data, and 3D skeleton point cloud data. It then uses a Transformer to fuse multimodal information and combines spatiotemporal features with an attention mechanism to construct the final model. Experimental results show that the average recognition rate of the model for police training actions is about 0.99, which is higher than the 0.91 of the Convolutional Neural Network Long Short Term Memory (CNN-LSTM) model, 0.93 of the Multi-Scale Temporary Feature Fusion Graph Convolutional Networks (MSTF-GCN) model, and 0.96 of the Channel-wise Topology Refinement Graph Convolution (CTR-GCN) model. The recognition rate of rescue training actions is 99.6%, higher than the 86.2% of the CNN-LSTM model, 93.3% of the MSTF-GCN model, and 91.8% of the CTR-GCN model. After 150 iterations, the model stabilizes with a training loss value of 0.35. These results demonstrate that the fusion of multimodal perception data effectively addresses the limitations of single-modal input, and the deep mining of spatiotemporal features improves both the accuracy and stability of motion trajectory recognition. This study provides technical support for intelligent training evaluation systems and helps promote the transformation of police training from experience-based to data-driven and precise approaches.</p>