Research On Object Matching For Disjoint Camera Views Based On Optimizing Distance Metric Learning

In the field of smart video surveillance, object matching for non-overlapping view between multiple cameras becomes an important and challenging problem, due to the changeable environment, different camera view, illumination change and other factors. Frome the viewpoint of Machine Learning, the thesis conducts exploration on this problem, combining appearance features and distance meteic learning optimization. The main work and innovations are as follows:(1)A correlation model between features and metric learning algorithms is proposed. By applying feature descriptors and metric learning algorithms, it not only looks for descriptors which having a significant degree of differentiation, but also analyzes nature relationship between the two and builds a correlation model. The model points out the most discriminative feature type under each metric learning algorithm. It can provide as a reference for object matching based on multiple features.(2) An adaptive online object matching algorithm which consists of original offline training, online matching and metric updating stages is proposed. For offline training and online matching model, the model parameter is fixed. Adaptive adjustment of metric matrix relies on the current state of samples which correctly matching. Moreover, the thesis puts forward a concept of color significant degree to help with weighting color histogram. The proposed method solves the disadvantages of common offline training & online matching model well. That is common model cannot be updated to incorporate the information carried by the new samples, which leads to seriously accuracy demotion.(3)Inspired by bidirectional ranking algorithm, a re-rank distance metric learning matching algorithm which improves efficiency of searching matched target is proposed. By conducting a forward and backward query, content similarity and context similarity are computed and later used for generating final similarity score and the new better rank list. In this way, target object sample has higher probability to appear in the k-nearest neighbors, which improves searching efficiency from given rank list. It solves the problem of unidirectional matching is sometimes unreliable.Experiments and simulations are conducted, and results show that our algorithms exceed common methods in matching accuracy. It verifies the effectiveness of algorithms in this thesis.The last portion gives a summary to the work of the whole thesis, and looks forward to the subsequent research.