A human motion behavior recognition method was proposed based on the DSC-BiGRU-Att Siamese networks, to address the challenge of achieving higher accuracy in human motion behavior recognition using fewer sensors. The method aimed to achieve accurate recognition of human motion behavior through deep learning network models using acceleration sensors. The pairs of pre-processed acceleration sensor data samples were inputted, based on the multi-classification task objective of human motion behavior, with Siamese networks as the basic framework. The Siamese networks consisted of two deeply separable convolutional (DSC) networks with identical structures and shared parameters, a bidirectional gated recurrent unit (BiGRU), and an attention mechanism (Att). The DSC network extracted spatial features from the motion behavior signal, while the BiGRU extracted temporal features. The attention mechanism was introduced to enhance the extracted features and ultimately facilitated human motion behavior recognition. Real-world datasets were utilized to validate the proposed DSC-BiGRU-Att Siamese networks. Experimental results demonstrated that the proposed algorithm achieved favorable behavior recognition results despite a small sample size, with an accuracy rate of 98.89% attained on the self-collected acceleration dataset.