Research On Human Action Detection And Recognition Algorithm For Video

Spatio-temporal action detection is a technology developed based on action recognition.Its purpose is to realize the classification recognition and space-time positioning of human action in video.The task requirements are oriented to real scenes and have high application value in real-life scenarios.Since video-oriented action detection and image-oriented target detection have great similarity in task mode,most of the current space-time action detection algorithms are designed with reference to the classical method of target detection in images.However,this kind of algorithm often does not detect the multi-scale action well when processing video or it takes a lot of computing resources and time,and has little application research in real-time online action detection.In view of the above problems,this paper mainly conducts the following research according to different application scenarios:1)For the video-oriented offline action detection task,this paper proposes a multi-stage human action recognition algorithm based on 3D convolution.The algorithm refers to the flow of R-CNN algorithm and is applied to the action detection task in video.A multi-scale action detection solution is proposed.Firstly,the Faster RCNN model is used for human body detection.Then,in order to obtain accurate and continuous human body sequence,a human body sequence connection algorithm is designed.Finally,in the action recognition process.Different from the C3D model input fixed-size video clip,this paper proposes a 3D convolutional action recognition model,which can input video frames of different sizes to achieve multi-scale recognition detection of human targets.2)In order to realize online real-time human action detection,this paper proposes a real-time human action detection model based on dual-stream YOLO.The algorithm relies on the framework of a typical dual-flow model and innovatively applies the YOLO model to the dual-flow network for action detection..Firstly,the YOLO v3 model is trained by video frame to realize the detection of moving targets in single-frame images.Secondly,a dual-stream fusion scheme based on RGB model and optical flow model is proposed to improve the robustness of the detection algorithm.Finally,according to the fusion Feature,this paper designs an online action pipeline generation algorithm to realize online detection of actions.By testing on the UCF dataset,the experimental results show that the two algorithms designed in this paper can complete the spatio-temporal action detection task well and achieve satisfactory results.Among them,the multi-stage human motion detection model based on 3D convolution is mainly applied to human action detection in offline tasks,which achieves high detection accuracy,and the real-time human action detection model based on dual-stream YOLO is mainly applied to real-time online action scenes.In this paper,different action detection technologies are provided for different application scenarios.