Ice and snow sports behavior recognition based on multi-scale features and improved CBAM

Accurately identifying and correcting erroneous sports behaviors of athletes or beginners in ice and snow sports can improve the training quality. However, ice and snow sports scenes often have complex motion backgrounds, and the behavioral features during motion are difficult to extract, which affects the recognition accuracy. In order to solve the feature extraction in ice and snow sports behavior recognition, a behavior recognition model based on multi-scale features and improved convolutional block attention module is proposed. The model first utilizes multi-scale features to obtain multi-level features from the collected ice and snow motion images, ensuring that features of different scales in the images can be effectively captured. Then, one-dimensional convolution and spatial random pooling layers are introduced to improve the convolutional attention module, thereby constructing a behavior recognition model. The accuracy of the proposed model in the Ski-Pose dataset was 98.3%, which was 8.2% and 13.7% higher than other recognition models, indicating an obvious gap. The accuracy and F1 value were 89.5% and 91.2%, respectively, and the recognition rate for small targets reached 80%, which verified the effectiveness of the model. The research provides new technological support for intelligent monitoring and analysis systems for ice and snow sports.