Research On Object Detection And Action Recognition Based On Video Stream

The rapid development of video processing technology is an important product of artificial intelligence at present.The detection of objects in video stream and the understanding of human activities have also attracted much attention.Traditional video task processing needs to construct feature model artificially.With the construction of complex features of tasks becoming more and more difficult,and the computational cost is getting higher and higher.With the complexity of tasks,the construction of features becomes much more difficult,and the computational cost becomes much higher.In this thesis,specific network structures are designed for video object detection and video action recognition to improve the detection efficiency and practicability of the network.For video object detection,a video object detection algorithm based on global vi-sion analysis is proposed.On the basis of the original flow-guided feature aggregation algorithm,more attention is paid to finding a compromise strategy between detection accuracy and running time.Firstly,using the idea of global vision analysis,the per-ceptual hashing algorithm is used to calculate the global visual similarity at both ends of the frame segment before aggregating multi-frame features,judging the correlation of image information in the current local frame segment.Secondly,continuous frames are used as input to further utilize the temporal information of the video,aggregate the features of adjacent frames on the motion path into the features of the current frame in order to better express the video features.Experiments on Image VID dataset show that the accuracy and speed of object detection in videos are improved after global visual analysis.For video action recognition,a video capsule network with attention mechanisms is proposed.On the basis of the video capsule network,the attention mechanism" squeeze-excitation" module adapted to the capsule network is embedded.By modeling the in-terdependence between capsule types,the ability of capturing action types of the whole network is enhanced,improving the performance of the network.The network can adaptively assign attention weight to different types of capsule features,selectively em-phasizing effective information features or suppressing useless ones by learning global information.Experiments on UCF-101(24 classes),J-HMDB and UCF-Sports action recognition datasets show that the proposed method achieves state-of-the-art perfor-mance.Experiments of the additional parameters show that the design is a lightweight gating mechanism,which could effectively model capsule type-wise relationships.An-alyzing the effect of attention mechanism on action recognition,verifying the effective-ness of the method.