Research On Behavior Recognition Algorithm Based On Deep Learning

Human action recognition is a key issue in machine vision,which has important ap-plication values in the fields of intelligent security,intelligent pension,human-computer interaction etc.Traditional methods of human behavior recognition have many shortcom-ings,such as the need for people to wear various sensors,the recognition of the action category is limited,the accuracy is not high.With the application of deeplearning in this field and the development of large-scale data sets,human action recognition technology has made rapid development.But in the real life scene,the complex and diverse back-ground and dynamic environment such as crowd gathering make the current method still unable to adapt to the diverse scene.In view of the existing problems,this paper studies the behavior recognition technology based on deep-learning algorithm as follows:Aiming at the problems of low recognition accuracy and poor adaptability to back-ground transformation of current human behavior recognition algorithms in real life scenes,we proposed a human action recognition method based on object detection.This paper analyzes the advantages and disadvantages of the classical object detection algo-rithm in terms of accuracy and speed,as well as the related theoretical principles.Based on the object detection network model YOLOv3,a single class human body detector is trained to improve the average accuracy and inference speed of the detection module.This paper analyzes the current action recognition network model with good effect,and uses the I3D model structure as the basic network of the action recognition module.Combined with human body detector and I3D network model,preprocess each frame of image through the information of human body position and existence state obtained by the detector,obtain the region of interest centered on human body,remove the background,and improve the recognition accuracy of behavior recognition algorithm in different backgroundIn the more complex multi-object dynamic scene,only relying on object detection can't realize the recognition of specific object behavior.We analyze the principle of multi-object tracking algorithm based on object detection,DeepSort,and replace the detector in the original method with YOLOv3,which has better performance,to improve the performance of tracking algorithm.Combined with multi-target tracking algorithm and behavior recognition algorithm,the object in the first frame is determined as the target to be identified,the tracker state is initialized,and the cost matrix is constructed based on the vector extracted from convolution neural network according to the boundary frame information of human body and the shallow features.The data association between frames and the corresponding object identity matching are realized through Hungarian algorithm,to identify the behavior of the target,and to recognize the behavior of the specific target in the multi-objective scene.At the end of this paper,we introduce the test dataset recorded in the actual scene,and on this basis,we test the method proposed above,and build a dangerous behavior detection system based on the Turtlebot robot to verify the effectiveness of the algorithm in different scenarios.