Multi-body Continuous Pose Recognition Method based on Deep Learning

Fangjuan Xie

Department of Physics and Electronic Information, Nanchang Normal University, Nanchang, 330032, China 

Abstract: At present, the recognition of human body posture is mainly carried out by collecting posture im-ages and establishing human body model to extract the key features of posture. Due to the limited selection of key points, it can only identify individual individuals, and the accuracy and efficiency of recognition is low. In order to improve the above defects, a multi-body continuous pose recognition method based on deep learning will be studied. The continuous pose of human body was collected by using Open pose and sensor equipment, and the sample set of human body pose was made. A 6-DOF motion model based on Twist description was introduced to analyze human posture and detect the dense key points of human body. The DCLSTM network model was established by combining CNN and RNN. By processing and classifying the sample data and key points, the real-time recognition of multi-body continuous posture was realized. The experimental results show that the proposed attitude recognition method is faster, and the overall recognition accuracy can reach 95.8%, which is better than the contrast method.

Keywords: Deep learning; Multi-body continuous posture; Real-time identification; Attitude recognition; Dense key points; CNN network; RNN network