Research Of Appearance Feature And Distance Metric Learning Based Moving Object Matching Across Disjoint Camera Views

Intelligent surveillance system based on multi-camera is attracting more and more widelyattention in real life with development of surveillance system. Object matching includesoverlapping view and non-overlapping view. Because non-overlapping view’s large monitoringrange, in this paper we will concentrate on object matching of non-overlapping view. Innon-overlapping view, the same objectives in terms of time and space is not continuous, so objectmatching is more challenging in non-overlapping view than in single camera and overlap objectmatching. In this paper to solve above problems, we studied the object matching algorithm ofnon-overlapping view. Innovation of this paper is as follows:(1) Colour feature is the commonly used appearance feature; it’s sensitive to the brightness ofthe scene change. Scene brightness varies between disjoint camera views. Brightness transferfunction can modify the changes of scene brightness. In this paper, we proposed a method ofmatching object to solve the situation that one scene have multiple brightness areas innon-overlapping view which uses mixtures of probabilistic principal component analysis to estimatesubspace of brightness transfer function.(2) Colour feature don’t include spatial distribution information, so different objects can havethe same colour feature. In this paper, we use color histogram、major color spatial histogram andhistogram of oriented gradient to improve robustness of object matching. And due to the dimensionof color space of major color spatial histogram is too high; the computing complexity is also high.Through reducing dimensionality of each color components, we reduced the dimension of colorspace and the complexity of computing.(3) Because of a deep semantic gap between low-level features and high-level concepts, sousing different way to calculate distance can get different result of object matching. In this paper,we use distance metric learning algorithm to learn a way of calculating distance of feature vector.Distance metric learning algorithm can give different weights for each dimension to highlight thegood performance feature. And on this basis, we proposed metric integrated algorithm based oninformation entropy.(4) Based on the above research, we put forward the appearance and metric learning basednon-overlapping comprehensive object matching algorithm. In the offline stage, we calculatebrightness transfer function, distance metric matrix, feature weighting; in the online stage, we extract fecture vector of object, and calculate the distance of feature vector to achieve objectmatching.This paper gives detailed experimental methods and the analysis of results, and verifies thecorrectness of the work. Finally, we get conclusion of this paper, and prospect the further researchwork.