文章摘要

文章信息/Info

Title:: Research on abnormal gait recognition based on efficient spatiotemporal graph convolution

作者:: 尹梓名1; 姜艺1; 喻洪流1; 2; 3△; 单新颖4; 于龚瑶1; 傅宇栋1; 罗军5; （1. 上海理工大学健康科学与工程学院,上海 200093；2.上海康复器械工程技术研究中心, 上海 200093；3.民政部神经功能信息与康复工程重点实验室，上海 200093;4. 国家康复辅具研究中心,北京 100176；5. 南昌大学第二附属医院康复科，南昌 330000）

Author(s):: YIN Ziming1; JIANG Yi1; YU Hongliu1; 2; 3; SHAN Xinying4; YU Gongyao1; FU Yudong1; LUO Jun5; （1. School of Health Science and Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China；2. Shanghai Engineering Research Center of Assistive Devices, Shanghai 200093; 3.Key Laboratory of Neural Function Information and Rehabilitation Engineering of the Ministry of Civil Affairs, Shanghai 200093; 4. National Research Center for Rehabilitation Technical Aids, Beijing 100176, China; 5. Department of Rehabilitation Medicine, The Second Affiliated Hospital of Nanchang University, Nanchang 330000, China）

Keywords:: Graph convolution network; Attention mechanism; Abnormal gait; Depth camera; Human skeleton

摘要:: 为实现人体异常步态的自动识别，本研究提出一种基于高效时空图卷积神经网络的异常步态识别算法，使用两个Kinect深度相机传感器提取人体三维骨架数据，基于时空图卷积采用早期多分支融合策略，生成关节、运动和骨骼三类特征。经特征融合后，使用两个时空图卷积块作为主流网络进行训练，结合时空关节注意力机制增强模型鉴别能力。在两个公开数据集上测试分别取得了99.37%和96.10%的平均准确率，实验结果高于其他基于骨架的图卷积神经网络方法。本研究提出的高效时空图卷积网络能有效鉴别异常步态，有助于异常步态的早发现、早诊断和早治疗。

Abstract:: In order to realize the automatic recognition of abnormal human gait，we proposed an abnormal gait recognition algorithm based on efficient spatiotemporal graph convolution neural network. Using the human 3D skeleton data extracted by two Kinect depth camera sensors, an early multi-branch fusion strategy was adopted on the basis of spatiotemporal graph convolution to generate three types of features of joints, motion and bones. Two spatiotemporal graph convolutions were used after feature fusion, and the block was trained as the mainstream network, the spatiotemporal joint attention mechanism was combined to enhance the model discrimination ability. The average accuracy of 99.37% and 96.10% were achieved on the two open datasets, respectively. The experimental results were higher than the other skeleton based graph convolution neural network methods. The efficient spatiotemporal graph convolution network can effectively identify abnormal gait, which can help early detection, diagnosis and treatment of abnormal gait.