Research On Retrieving Method For Abnormal Events Of Surveillance Videos Based On Temporal Action Detection

In recent years,the importance of surveillance video data in urban security work has become increasingly significant,and has important research value.The surveillance video records a variety of real abnormal events,including human abnormal behaviors(such as illegal and criminal actions)and abnormal phenomena(such as construction fire).These abnormal events are indispensably important information in security work such as reconnaissance.However,unlike normal events,the probability of anomalous events is small,and it is very difficult to find the required information of anomalous events in massive surveillance video data.Traditional retrieval methods cannot meet the needs of analyzing abnormal events,and when processing massive surveillance video data,the retrieval efficiency and retrieval performance are still insufficient.Therefore,this article takes the abnormal events as the research object,researches and improves the key parts of the retrieval technology,and realizes the retrieval method of quickly and effectively retrieving the abnormal video fragments of the surveillance video related to the image content based on the query image.The main research contents of this article are as follows:1.The surveillance videos have a lot of redundant information and it is difficult to obtain information of abnormal event in it.Aiming at the above problem,this paper studies the temporal action detection technology and proposes an abnormal event detection algorithm based on the temporal action detection technology.The abnormal event detection algorithm is used to extract abnormal event fragments from the original surveillance video and remove a lot of redundant information,so it can obtain the video database of abnormal events which is much smaller than the original surveillance video data volume.It is convenient for subsequent retrieval.Aiming at the problem of low accuracy of the current multi-stage temporal action detection algorithm,an improvement scheme is proposed: the C3D(3D Conv Nets)network and multi-instance learning method are used to construct a supplementary generator of abnormal event fragments and modify the structure of the classification network.Experiments show that the algorithm proposed in this paper can effectively improve detection performance and accurately obtain abnormal event fragments in surveillance video.2.The surveillance video has a large amount of data,the query data and database data have asymmetry.Aiming at the problems,this paper proposes an efficient video retrieval algorithm.Feature extraction,index construction and distance measurement are performed on the abnormal event database.First,the SCFV(Scalable Compressed Fisher Vector)method is used to generate video feature vector which occupies less memory and is simple to calculate.Then the database technology is used to construct a multi-level index structure for the video feature vectors,so that the retrieval method can still obtain better retrieval results in less time when processing massive high-dimensional surveillance video data.Then,considering the asymmetric relationship between the query data and the database data,an asymmetric comparison algorithm is used to implement the distance measurement between the feature vectors.The above methods are verified through multiple sets of experiments,and it is confirmed that the retrieval algorithm proposed in this paper effectively solves the above problems,has a good performance in terms of efficiency and accuracy,and meets the needs of practical applications.