Indoor Person Behavior Recognition Based On Deep Learning And Knowledge Reasoning

With the development of computer technology,human behavior recognition has been greatly applied in the fields of video intelligent monitoring,virtual reality and film production.This paper "Indoor Person Behavior Recognition Based on Deep Learning and Knowledge Reasoning" aims to complete the comprehensive recognition of personnel behavior in the scene through a variety of technologies,so that the recognition results are interpretable.The classic human behavior recognition method is to build an end-to-end neural network model,which is a traditional image classification problem by inputting images into the network and then classifying the behavior categories according to the characteristics extracted by the network.In the research scheme of this paper.Firstly,the human key points and scene target detection are completed through the constructed deep neural network,and the feature triples in the scene that used for behavior reasoning are obtained;Secondly,inference rules and behavior recognition knowledge graph are constructed according to the spatial and semantic relationships between human behavior and key points and scene objects;Finally,combined with the behavior feature triples in the scene and inference rules to complete the reasoning recognition of behavior.The main research work includes:1.A human key points detection model based on Dense Residual Steps Network was constructed.The advantages and disadvantages of two basic network structures,Dense-Net and Residual-Net,are analyzed as the backbone network.Based on these two network ideas,a convolutional wide receptive field and real-time detection Dense Residual Steps Network(DRSN)is proposed to realize the detection of human key points,which improves the ability of extracting for the spatial information and the global features for the input image.At the same time,the FReLU two-dimensional convolution activation function is analyzed.The improved FReLU activation function is used to replace the original activation function so as to reduce the model parameters of the backbone network,which accelerates the prediction speed of the model and improve the positioning accuracy of key points.2.A triplet recognition method of behavior characteristics based on target detection and human key points detection is studied.Using the original MSCOCO data set build our data set and using the YOLO v5 model to complete the detection of objects such as beds,chairs,cup,water bottles,computers,keyboards,mouse,books,etc.At the same time,combined with the positioning of human key points and scene target detection results,the visual space relationship between human body parts and scene objects is obtained according to the pixel space distance and threshold discrimination method.Finally,all feature triples that can be used for behavioral reasoning in the scene are obtained.3.The knowledge graph of behavior recognition is constructed and the knowledge inference method based on rule inference is given.Firstly,the priori knowledge of potential spatial and semantic relationships between human behavior and scene objects and human body parts is analyzed,and the knowledge graph of behavior recognition is constructed.Secondly,the inference rules corresponding to various behaviors are constructed based on the behavior recognition knowledge graph.Finally,according to the behavior feature triples detected in the scene,combined with the inference rules to complete the reasoning recognition of the human behavior in the indoor.By constructing an indoor experimental scene,the model and technical scheme designed in this paper are experimentally verified.Under the condition of high accuracy,the location and classification of human key points can achieve super real-time effect.Combined with scene target detection,human key points positioning and knowledge reasoning,the comprehensive recognition accuracy rate of human behavior reasoning reaches 94.7%,which enhances the interpretability of recognition results under the condition that certain recognition effects are met.