Research On Key Technologies Of Human Action Recognition Based On 3D Skeleton

The advances in microelectronic technology,network technology and sensor technology provide unprecedented convenience for real-time continuous capture and collection of human activity states.The perception of sensors that originate from different modal and types is essentially a true record of human behavior.The analysis of these behavioral data can help us to comprehensively understand the human behavior pattern and its transition process.It is even impossible to find the unknown behavior pattern.Thus it has wide application in the fields of safety monitoring,scene recognition,intelligent assistant,human-computer and so on.This paper mainly studies the human behavior recognition based on 3D skeleton information.It proposes two kinds of recognition methods,which are the behavior recognition based on mixed joints and the behavior recognition based on multi-modal features.The main research work in this paper has the following three aspects:First,this paper analyzes the research background and significance of human behavior recognition.It summarizes the research status of human behavior recognition from three kinds of data types,such as RGB data,depth data and skeletal data.What's more,it analyzes its advantages and disadvantages of different data types and proposes the research methods according to the research objectives.Secondly,in view of the shortcomings such as low recognition efficiency and strong environmental disturbance in traditional behavior recognition methods,this paper presents a new kind of skeletal descriptor with mixed joint features.This descriptor integrates both kinetic and potential energy on the basis of skeletal information,making the feature description more abundant.At the same time,this paper chooses the LSTM(Long short term memory)neural network as a classifier to better study the characteristics of mixed joints and make important contributions to the classification of accurate behavior.Thirdly,this paper finds that accurate identification for complex behavior based on the hybrid joint feature can not be made.So a behavior recognition algorithm based on multi-modal feature is proposed.The multimodal features,which include hybrid joint features and HDMM features,combine skeletal information with deep image information.The advantage is that when any of the two types of information has serious loss or noise problems during the acquisition process,another type of information can be used as a supplement,which has a good effect on human behavior recognition.