基于Kinect的双流时空卷积人体行为识别技术

Kinect-based dual-stream spatiotemporal convolution human behavior recognition technology

针对原有双流时空卷积网络模型中网络深度不足,从而导致人体行为识别结果偏低的问题,针对该网络模型进行改进,且融入Kinect骨骼序列数据.对于输入数据,通过Kinect相机对人体动作转化为骨骼序列;改进双流卷积网络模型是在原有的模型框架下,用RestNet-50网络结构替代原VGG-16网络结构,再对数据进行一系列处理.在HMDB-51和UCF-101两个公开数据集上进行模型的训练和验证,其识别结果分别为70.8%和91.4%,通过对比结果表明,本文提出的改进双流卷积网络融合Kinect骨骼数据能够有效提升人体行为识别的正确率.

In response to the problem of insufficient network depth in the original dual stream spatiotemporal convolutional network model,which leads to low recognition results of human behavior,this paper improved the network model and incorporates Kinect bone sequence data.For input data,this paper used Kinect cameras to convert human movements into skeletal sequences;the improved dual stream convolutional network model replaced the original VGG-16 network structure with the RestNet-50 network structure under the original model framework,and then performed a series of data processing.The model was trained and validated on two publicly available datasets,HMDB-51 and UCF-101,and the recognition results were 70.8%and 91.4%,respectively.The comparison results showed that the improved dual flow convolutional network proposed in this paper for fusing Kinect bone data can effectively improve the accuracy of human behavior recognition.