Abnormal Behavior Detection Using Sparse Coding and HOG3D Descriptor

In this paper,an abnormality behavior detect method based on sparse coding is proposed and the HOG3D descriptor is utilized to capture appearance and motion information of the surveillance videos.Firstly,a set of training data are extracted from normal events.And then,K-SVD method is utilized to construct the dictionary atoms such that each normal member attains the best representation under the strict sparsity constraints.In the process of sparse coding,by taking a video session as a sample,we introduce the relative sparse reconstruction error over the normal dictionary to measure the level of normal of the testing sample.When the relative sparse reconstruction error is positive,the sample would be judged as abnormal.The proposed method is tested via UMN database,Weizmann database and real world surveillance videos,which show that the proposed method can reliably detect the unusual events in the video sequence.