During the process of assisted walking or gait rehabilitation training, it is essential to recognize abnormal behaviors accurately in human-computer interaction on the basis of following user′s gait closely. This article proposes a non-contact recognition method which has advantages of universality, robustness, and convenience for multimodal walking intention. It can control the robot flexibly and accurately recognize various gaits. Firstly, the structure, functions, and kinematics models of the walking assist robot and the gait rehabilitation training robot are introduced, and an embedded airborne gait recognition system is established. It can accurately describe the gait and changing rule. Secondly, to effectively solve the problem of mark point loss, a new extended set membership filter is proposed to estimate the knee angle. Finally, a compliance control method based on walking speed compensation is established by combining with gait information. Experimental results show that the proposed method could effectively overcome the loss of marker points, identify the normal gait accurately in the interaction process, and flexibly control the robot movement. Meanwhile, it can effectively recognize the falling and drag-to-drop gait. The recognition rates are 91. 3% and 89. 3% , respectively. The non-contact walking intention recognition method can be applied to walkers with similar structures and their daily walking assistance or rehabilitation training