Abstract: Aiming at the problems of poor detection accuracy of current human motion recognition algorithms and the diversity of experimental scenes, a new human motion recognition method based on hybrid convolutional neural network-hidden Markov model (CNN-HMM) is proposed. In order to verify the effectiveness of the method, we establish three sets of human rehabilitation training motion models including one standard motion posture and five non-standard motion postures for leg-lifting, squat and hip bridge, respectively. The experimental data are obtained by the wearable inertial motion capture system, Perception Neuron 2.0 (PN2.0). Finally, the performance of the proposed method is evaluated in terms of accuracy, sensitivity and specificity. Three groups of the experimental results show that the proposed method can distinguish the six different motion gestures with a high average recognition rate of 97.00%, which is 5.78% higher than the single CNN method.