Activity Correlation Analysis And Person Re-identification Of Intelligent Multi-camera Video Surveillance System

With the development of the society, people demand increasingly strong regionalmonitoring system for public safety. In recent years, the rapid developments ofcomputer technology effectively promote the basic research and applicationpromotion of intelligent multi-camera video surveillance system. As two keyproblems of the intelligent video surveillance system, activity correlation analysis andperson re-identification attract vast number of scientific research workers.Activity correlation analysis refers to analysis of the correlation of each cameraneutron area coverage and time delay. In practical application, we can get the cameratopology, for subsequent person re-identification problem again. While personre-identification refers to regarding the video monitoring in pedestrian detectionresults under a single camera as an input, then match it with the pedestrian detectionresults under the other cameras. This has greatly apply in the security monitoring.In practical application of public security area, the key point of activitycorrelation analysis is that it’s difficult to extract both effective and stable features ofactivity generated from the crowds and vehicles, which is the basis of correlationanalysis; only the stable feature has actual significance to study the useful correlation.In order to solve this problem, this paper proposes a method based on slow featureanalysis (SFA) and applies it to the low-level feature of activity in order to obtainmore meaningful and stable high-lever feature representation.While the difficulty of person re-identification lies in how to choose an optimalmeasure to match pedestrians of different cameras. Because of changes from theillumination, angle differences under different cameras, and person’s posture, thesame persons from different cameras will have very big difference in visualcharacteristics. In order to solve this problem, this paper proposes a sparse codingbased measure. Through the training the relative distances of pairwise samples from same pedestrian captured in different cameras to obtain a dictionary, systemicdifferences of different cameras can be eliminated at the same time. Then relativedistances pairs of probe samples and gallery samples will be reconstructed with thelearned dictionary, and the reconstruction error is a criterion to measure the result ofmatching. The smallest error of pairs will be regard as matched pairs.In order to verify the effectiveness of the two methods proposed in this paper,they are investigated on the video data in a fork road and three public personre-identification databases (i-LIDS, VIPeR, and GRID) for experimental verification,and compared with the state-of-the-art methods in individual field respectively. Theexperimental results show the effectiveness and superiority of the two proposedalgorithms.