| With the continuous development of Internet multimedia technology,the amount of video data on the network has increased dramatically.Therefore,there is a huge need for fast and accurate content understanding and analysis of massive videos,and it has gradually become a research focus in the field of computer vision in recent years.Among them,video action detection is an important branch of research on video content understanding,the purpose of which is to detect the time and place of action occurrence from the uncropped video and identify the type of action.However,current video action detection algorithms suffer from low detection accuracy and long detection time.This article designs and analyzes an existing video action detection network by summarizing and analyzing related issues in the existing video content understanding field,including video action recognition,video action candidate frame generation,and video action detection.The main work and contributions of this article are as follows:(1)The video action classification network is integrated into the network structure of the video action candidate frame generation algorithm.Using network sharing and feature sharing,a branch of a classification score is predicted,and the action classification result is obtained at the same time,which reduces the time complexity of the algorithm.(2)Add embedding vectors and boundary compensation branches to the network.By predicting the embedding vector of each video unit,more accurate video action candidate frames can be matched.And predict boundary compensation to improve the accuracy of video detection boundary results.(3)The specific design and implementation of the video action detection algorithm in this paper.The Non-Local module algorithm is improved in the video feature extraction module.Finally,a more accurate output is obtained using the Greedy-NMS post-processing.The Anchor Free video action detection network in this article was tested on the THUMOS-2014 and ActivityNet-1.3 datasets.The experimental results show that the accuracy rate is improved by 0.84%and the time performance is increased by 18%.The improved Non-Local module and the Greedy-NMS post-processing,the module also improves the overall performance of the algorithm. |
