Research On Key Point Based Periodic Action Detection And Action Proposal Generation Method

Temporal action detection is an important subject in the field of computer vision.Based on this topic,many in-depth research and application can be done,from realtime applications such as extracting highlights from sports videos to automatic video subtitle description,intelligent monitoring,and intelligent security.Periodic action detection is helpful for basic visual tasks such as video tracking,and also for other advanced recognition tasks such as analyzing and identifying individuals based on periodic gait on individual.Based on the problem of temporal action detection,this dissertation respectively researches periodic temporal action detection and general purpose temporal action detection.The research work in this article will include the following two parts: 1.Research on the periodic temporal action detection method based on key point detection.This paper combines the classic method of key point detection OpenPose and the classification results of active areas in the video to construct a two path method to realize the analysis of human body's periodic action localization.In this paper,OpenPose is used to extract the joint point trajectory,and CNN is used to obtain the local features of the main active area of the individual.Then the two kinds of information are fused by the two threshold fusion method proposed in this paper.This paper uses the time intersection-over-Union(tIOU)of predicted localization and GroundTruth and mAP accuracy at different tIoU as metrics.The experimental data from real factory scene.The method proposed in this paper is applied to the sewing industry video to have a certain validity for localization of normal working human action and achieve high accuracy.2.Research on temporal action detection method of Boundary sensitive network based on improved loss function.In order to improve the accuracy of neural network in temporal action detection,this paper proposed a new improved loss function for BSN network,which increases the original loss function to limit the order of action procedure,and can better optimize the loss function from a more reasonable direction.The experimental results on public dataset activitynet1.3 show that the performance of the loss function is better and more reasonable than the original model loss function.