Surveillance video exceptional event detection method based on deep learning and dynamic clustering

Abstract

The invention relates to a surveillance video exceptional event detection method based on deep learning and dynamic clustering. In a characteristic extraction stage, a deep learning network PCA (Principal Component Analysis) Net is applied, a video is trained to learn a corresponding network filter, low-layer pixel optical flow characteristics are converted into high-layer semantic motion characteristics through a deep network, and meanwhile, motion areas in a video are screened to remove a spatial-temporal sampling block which only contains background information. In a characteristic modelingstage, a nonparametric model based on two-layer clustering is applied to carry out modeling of characteristic vector space, a vector opposite-direction combination method is adopted in a vector combination stage, finally, a K-means clustering algorithm is applied for clustering vectors in a dictionary set into one series of event clusters, and an exceptional event is judged according to Euclideandistance between a test vector and an event cluster central vector. By use of the method, characteristic vector offset caused by addition can be effectively avoided, and an exceptional event detection rate is improved.

Description

technical field [0001] The present invention relates to a monitoring video abnormal event detection method, in particular to a monitoring video abnormal event detection method based on deep learning and dynamic clustering. Background technique [0002] With the development of computer science and technology, the use of image processing, computer vision, machine learning and other technologies can break through the limitations of traditional video surveillance systems, and realize intelligent video analysis of video surveillance systems and active detection and real-time warning of abnormal events. Video surveillance applications in the security field are of great value. [0003] The abnormal event detection method in surveillance video is mainly divided into four basic steps: image preprocessing, basic event representation, building an abnormal detection model and judging abnormal events. Among them, the basic event representation is mainly divided into event representation...

AI Technical Summary

Problems solved by technology

[0006] 1. The algorithm uses MHOF features to describe the motion in the video. Although the description effect of artificially constructed features such as HOF and HOG is good, the applicability of various features in different video scenes is different. Changing the scene often needs to be changed at the same time. The features used are not suitable for abnormal event detection in multiple scenarios;

[0007] 2. The algorithm adopts a simple weighted addition method in the vector merging of the dictionary set, which will cause the value of the feature vector in the dictionary set to shift relative to the original value after a large number of vector updates, which will affect the final Detection has an impact;

[0008] 3. In this algorithm, the detection of low-frequency abnormal events is carried out by counting the frequency of occurrence of vectors in the dictionary set and calculating the frequency ratio of the corresponding codebook. However, in the feature extraction stage, dense sampling of the entire image is used. , so that when the video scene is a sparse scene, most of the feature vectors obtained by sampling are background information, so the frequency count value of the vector representing the background information in the dictionary set will be very large, making the corresponding codebook frequency ratio too high , causing the frequency of other motion events to be less than the judgment threshold, resulting in false detection

Images