Abstract: A recognition method for continuous gesture action sequences is proposed to overcome the limitations of existing research, which relies on manually segmented recorded data while ignoring the continuity of hand actions. By automatically segmenting sequences of arbitrary length, the method enables independent recognition of stable segments. During data collection, a strategy of free spatial angle variation is employed to enhance data diversity and model robustness, ensuring adaptability to different users and environments. To achieve automatic segmentation, a cosine similarity-based dual-threshold method is designed to effectively identify stable segments. Geometric features based on specific triangulation areas of point clouds are introduced, enhancing the representation of the hand action spatial structure and significantly improving recognition accuracy. On a self-built recognition dataset, the method achieves an accuracy rate of 97.09% when combined with the Time Series Transformer model. For the self-collected continuous gesture action sequence dataset, it achieves a 100% success rate in automatic segmentation of individual segments and a recognition accuracy of 97.92%.