Improved HOF Based Anomaly Evente Detection Algorithm

Abnormal event detection is a hot research topic in the field of computer vision.Abnormal event detection can be widely used in video surveillance,intelligent detection and other fields.Breakthrough of this problem can bring considerable social value and economic benefits,such as reducing the cost of human resources in monitoring,improving the timeliness and accuracy of video surveillance.Abnormal event detection is a binary classification problem.Sparse representation is a mainstream method in non-deep learning classification methods.Sparse representation can abstract the training sample vector into a sparse matrix,which can effectively restore the training sample vector linearly.Based on this feature,if the sparse matrix can be used to effectively restore a vector,it shows that the vector is the same type of vector as the training sample.On the contrary,if the sparse matrix can not restore a vector well,it shows that the vector is very different from the training sample.Starting with the improvement of optical flow calculation method of HOF,this paper improves the application of HOF in abnormal event detection.Computing optical flow field is a commonly used analysis method in video research.Optical flow field represents the instantaneous velocity of a moving object in space on the observed imaging plane.The HOF feature of optical flow histogram can be obtained by mapping and statistic the direction of optical flow field in a certain region.HOF is a feature vector with strong generality,which is used in pattern recognition,motion detection and other problems.At the same time,because of its generality,HOF has a lot of room for improvement when it is applied to a particular problem.This paper compares the performance of HOF in abnormal event detection by using different optical flow algorithms to calculate HOF features.To solve the problem of anomaly event detection,the calculation method of HOF is improved.In this paper,the HOF value of the characteristic point trajectory is used as the eigenvector.The HOF value extracted from the training video is used as the training sparse matrix.For the video to be detected,the HOF characteristics of its trajectory in a certain period of time are calculated,and then the fraction values of abnormal events in the video in that period are obtained by combining them with the sparse matrix obtained by training.