Research On Key Technologies Of Structral Analysis On Surveillance Video

With the development of society,surveillance cameras are now everywhere in life,and surveillance videos record varieties of information.Although the information expressed by video is more intuitive,the cost of video storage is high,and it is not easy to quickly retrieve the information,it cannot serve us better.Therefore,how to quickly analyze video content and save storage costs has become an urgent problem to be solved.Video structured analysis technology extracts rich information in video through image processing,content recognition and other methods,and semantically describes the content of video analysis,achieving the purpose of facilitating structured storage and quick retrieval of video information.Based on deep learning and image processing,this thesis studies some key technologies of video structured analysis.The main research contents are as follows:1.Based on the model lightweight theory and target detection algorithm,this thesis proposes a new lightweight target detection network.The network introduced a multi-scale fusion module and a spatial attention module to enhance the feature expression capabilities of the network,and achieved very competitive results on public data sets,while realizing real-time detection on the CPU server.2.Aiming at the problem of pedestrian attribute recognition,a new attention module is proposed to automatically extract local features related to attributes in the attention area;then,the accuracy of attribute prediction is improved by fusing global features.In addition,a loss function based on dynamic weights is proposed to balance the loss contribution of positive and negative samples,effectively alleviating the problem of unbalanced sample distribution and enhancing the generalization of the model.3.Combining target tracking,main color extraction and other technologies,a video structured analysis system is designed and implemented,which verifies the effectiveness of the proposed algorithm.