Research On Key Algorithms Of Human Behavior Recognition Based On Convolutional Neural Network

Human behavior recognition refers to the use of a specific device and detection algorithm to detect the motion state of the detected object.Human behavior recognition technology has broad application prospects and potential economic value in intelligent video surveillance,human-computer interaction and video content retrieval.How to make the machine quickly and accurately recognize all kinds of behavioral information issued by the human body and how to improve the sensitivity and ease of use of the machine are the focus of the research on human body recognition technology.The mainstream human behavior recognition methods include video-based methods,wearable device based methods and Wi Fi signal based methods.The algorithm of human behavior recognition technology based on video has high computing cost and poor universality.Moreover,this technology relies on camera equipment and has a high risk of privacy leakage.The human behavior recognition method based on wearable devices has its own invasions and low comfort.The human behavior recognition method based on Wi Fi signal does not require wearable devices,has low computational cost,and avoids the risk of privacy leakage.Therefore,it is a popular research direction to realize human behavior recognition at present.In human behavior recognition process,the existing classification algorithm is affected by extract the characteristics of the type and quantity is big,in order to omit artificial feature extraction steps,the paper proposed a method of using deep learning,make network model to study the characteristics of the data,so as to avoid the artificial extracting feature of recognition on the chance,realize the recognition of human behavior.In this paper,human behavior is divided into human movement and gait,and recognition methods are designed for human movement recognition and gait based identity recognition respectively,including data acquisition,data preprocessing,data set construction,training model and classification recognition.The main contents are as follows:(1)For human motion recognition,a shallow convolutional neural network model: Rec CNN is proposed.By integrating the batch regularization mechanism into the convolutional layer,the convergence speed of the algorithm can be accelerated,the learning efficiency of the convolutional layer can be improved,and the recognition accuracy can be improved.Thus,the classification and recognition of different actions of the same observation target can be achieved with low cost and high precision.(2)Aiming at human behavior recognition based on gait,v GGnet-16 is improved based on transfer learning idea and applied to gait identity recognition.The structure of the model is optimized and adjusted to better adapt to the amplitude characteristics of the signal,the convergence speed of the algorithm is increased,the recognition accuracy is improved,and the high-precision identity recognition based on gait is realized.Compared with similar models and traditional models in the industry,the recognition accuracy is improved by 3.22% to 5.57%.The results show that the model recognition accuracy decreases by 9.32% to 13.15% in the sparse environment and the multipath environment,and the model has good robustness.To sum up,this paper proposes an Rec CNN model for human movement recognition and an improved VGGnet-16 model for gait identification,which has certain application value and practical significance.