Ice and snow sports action recognition and danger warning based on neural networks algorithm

The goals of this study are to accurately identify the behaviors of ice and snow sports, overcome the shortcomings of conventional approaches in handling the intricate dynamics of these activities and ambient noise interference, and enable prompt hazard warning. In terms of technical design, the study collects ice and snow sports behavior data through smartphone sensors, introduces a new adaptive factor into the traditional Kalman filter to adjust the Kalman gain in real time, enhance the realism of state information estimation, and reduce external noise interference. Meanwhile, it proposes a 1D Chaos-ResNet network for sports behavior recognition, integrates the ResNet50 network into ChaosNet, and constructs the deepChaosNet model, which effectively alleviates the class imbalance problem and enhances the recognition ability for minority class data. This study uses multiple sensors to collect key ice and snow sports data at a sampling frequency of 100 Hz, constructs a dataset containing 5,000 action samples covering movements such as skiing and turning, and annotates them with professional coaches. The environmental perception system based on the extended Kalman filter achieves a tracking accuracy of 95.3% and a detection accuracy of 96.6% on smooth, obstacle-free snow pathways, as demonstrated by the findings. The system performs worse on snow pathways with obstacles, though, with a tracking accuracy of 92.6% and a detection accuracy of 93.6%. The deepChaosNet model performs excellently in action recognition tasks, with an overall accuracy of 94.5% and a recall rate of 95.2% for critical actions such as falls. The effectiveness of the proposed deepChaosNet model in solving the class imbalance problem provides new ideas for ice and snow sports safety protection and helps reduce accident risks.