深度视频下的人体动作识别研究

Research on Human Motion Recognition in Depth Video

目的基于RGB视频序列分类是实现人体动作识别的主要方式,但是RGB视频在记录人体动作的同时会清晰地保存人体的面部信息,为保护隐私,本文提出基于深度视频进行人体动作识别。方法利用公开数据集UTD-MHAD中27种深度视频形式的动作数据进行研究。首先,将深度视频序列进行预处理转化成运动历史图,通过伪彩色编码增强运动历史图的细节信息;其次,将经过伪彩色编码的运动历史图送入经过预训练的卷积神经网络提取运动历史图的深度特征向量;最后运用分类器进行分类。结果基于深度视频序列的人体动作识别方法在UTD-MHAD数据集上取得了90.02%的准确率,误差为1.8%。结论本文提出的基于深度视频序列的人体动作识别方法具有一定的有效性,可作为人体动作识别领域一种新型的无监督康复锻炼手段,有助于促进康复评定研究进一步标准化。

Objective Classification based on RGB video sequence is the main way to realize human motion recognition.However,RGB video can clearly preserve human facial information while recording human motion.In order to protect privacy,this paper proposes human motion recognition based on depth video.Methods The public UTD-MHAD dataset that contained 27 kinds of depth video motion data was used for research.Firstly,the depth video sequence were preprocessed and transformed into motion history map,and the detail information of motion history map were enhanced by pseudo color coding.Secondly,the pseudo color coded motion history map were sent to the pre-trained convolutional neural network(CNN)to extract the depth feature vector of the motion history map.Finally,the classifiers were used for classification.Results The human motion recognition method based on depth video sequence achieved 90.02%accuracy and 1.8%error in UTD-MHAD dataset.Conclusion The human motion recognition method based on depth video sequence method proposed in this paper has certain validity,and can be used as a new unsupervised rehabilitation exercise method in the field of human motion recognition,which is helpful to promote the further standardization of rehabilitation evaluation research.