Abstract
This study proposes a novel gesture recognition system based on LoRa technology, integrating advanced signal preprocessing, adaptive segmentation algorithms, and an improved SS-ResNet50 deep learning model. Through the combination of residual learning and dynamic convolution techniques, the SS-ResNet50 model significantly enhances the extraction capability of multi-scale gesture features, thereby augmenting the classification accuracy. To counter environmental noise and static interferences, an adaptive segmentation approach based on sliding window variance analysis is introduced in the research. This method effectively increases data diversity while preserving the specific components of gestures. Experimental outcomes indicate that the system exhibits strong robustness in cross-scenario and cross-device tests, with an average recognition accuracy of over 95% for six gestures. The system's low power consumption, long-distance communication, and strong anti-interference capabilities offer broad prospects for its application in complex environments, particularly in resource-constrained scenarios such as underground mine gesture monitoring and remote control in dynamic environments and other practical applications. This study demonstrates the feasibility of gesture recognition systems based on LoRa technology and provides a new solution for low-power, long-distance non-contact gesture recognition.