Abnormal Detection Method Of Multi-feature Video Based On Sparse Coding

Abnormal event detection in large-scale video is an important task in research and industrial applications,and has received much attention in recent years.Existing methods usually solve this problem by extracting local features and then learning the anomaly detection model in the training video samples.However,most of the previous methods used only hand-crafted visual features that were limited in terms of presentation capabilities in video.In this paper,an anomaly detection method of multi-feature video based on sparse coding is proposed,which can generate effective feature information and can effectively detect the active mode of abnormal events in video.The method is mainly composed of four parts,including dictionary learning,deep three-dimensional convolution network,construction of multi-level similar tree and triple loss.First,hand-crafted visual features,such as 3D gradient features,are utilized as input to sparse coding to guide unsupervised feature learning.Then,the result of the sparse coding is constructed into a multi-level similarity tree,and the multi-level similarity tree will be used as the input of the deep three-dimensional convolutional network training.Secondly,it is necessary to construct the triple loss of multi-level similar structure according to the multi-level similarity tree,which is used as the supervised signal of the training depth three-dimensional convolution network.The deep three-dimensional convolutional network can generate features in the time dimension and the spatial dimension,which are simply referred to as spatiotemporal features.The spatiotemporal feature is a good representation of the rich information in the video.The newly generated spatiotemporal features can be sparsely encoded again.Robust and rich feature representations can be obtained by iterative updating between sparse coding and unsupervised feature learning.Finally,the sparse reconstruction error is applied to predict the abnormality score of each test input,and the abnormal event in the video is determined.This paper verifies the effectiveness of the proposed method on the challenging anomaly datasets,such as Avenue datasets,UCSD datasets,and Subway datasets,and compares them with advanced anomaly detection methods.It is very competitive.Finally,based on the video anomaly detection method proposed above,this paper proposes a n anomaly detection system construction scheme of multi-feature video based on sparse coding,and gives a detailed description of the overall architecture,system flow and system scheme.The system is very scalable and facilitates the development of anomalous event detection in video.