Research On Daily Human Activity Recognition Technology Combined With Object Detection

In recent years,the development of artificial intelligence technology is in full swing,which has touched all aspects of human life and has dramatically changed people's production and lifestyles.Human action recognition technology aims to analyze human behavior by computer and is an important research direction in artificial intelligence.It has broad application prospects in health monitoring,security monitoring,humancomputer interaction,virtual reality,etc.For example,human action recognition technology can monitor falls of the elderly or dangerous behaviors in a public place to respond in time when an accident occurs.This subject focuses on image-based human action recognition,studying two aspects of “how to locate and distinguish multiple human targets in an image” and “how to recognize the action of a single human target”.The thesis first proposes an object detection algorithm based on feature fusion of neighboring scales to locate human targets in an image.Then,based on the middle layer features and detection results of the object detection model,respectively proposes an action recognition algorithm based on object detection and an action recognition algorithm combining object detection and human image features.The research contents are as follows:1)By analyzing the advantages and disadvantages of some classic feature pyramid networks,proposing a feature pyramid network based on neighboring scale fusion.Based on Retina Net and ATSS,different feature extraction backbone networks and classification and location networks are used to build the object detection model.On the MS COCO 2017 dataset,the detection accuracy and inference speeds are 44.5% and 10.1 FPS.By reducing the resolution of input images,the inference speed can increase to 15.7FPS,and the corresponding detection accuracy is 43.3%.2)By reusing the middle layer features generated by the object detection model,proposing a human action recognition algorithm based on object detection.On the VOC 2012 and Stanford40 action recognition datasets containing the middle layer features of the object detection model,the Top1 classification accuracy is 85.85% and 77.33%,respectively,and the corresponding Top3 classification accuracy is 97.08% and90.77%.It is verified that the middle layer features of the object detection model can help action recognition.3)The thesis studies the feature extraction of the human image with action recognition as the optimization goal.Based on the human image features and the middle layer features of the object detection model,proposing an action recognition algorithm combining object detection and human image features.On the VOC 2012 and Stanford40 action recognition datasets,the Top1 classification accuracy is 91.13% and87.68%,respectively,and the corresponding Top3 classification accuracy is 98.34% and 96.07%.It verifies the validity of the re-extracted human image features and also confirms that the middle layer features of the object detection model can help human action recognition.