Research On Multi-Region Object Tracking Algorithm With Flexibility Of Illumination Change

In recent years, object tracking is an important task within computer vision with a wide range of applications, for example video surveillance, activity analysis, classification and recognition from motion, human-computer interfaces and so on. The proliferation of high-powered computers, high-capacity memory and high-quality video cameras, and the increasing need for automated video analysis has made object tracking have an amazing application prospect.This subject has two purposes. Firstly, the basic framework and the different classification methods of object tacking modules are studied. Secondly, although fragments-based tracking owns the ability to handle occlusion, it doesn’t have the flexibility of illumination change. The major contributions of this task consist of four aspects:Firstly, the framework of target tracking and the different classification methods of each module are discussed. In details, and several classical algorithms and their applications environments are introduced well.Secondly, the thesis analyzes a classical robust fragments-based tracking algorithm, in which the templates target is expressed by multiple image regions, the regions are arbitrary and do not refer to any particular object models. Every region votes on the candidate positions of the object in the current frame, by comparing the corresponding image region histogram with its histograms. Then the best matching point is obtained through the voting matrix. The key of the algorithm is to use the integral histogram structure. It allows to extract histograms of multiple rectangular regions in a very efficient manner. The method overcomes several difficulties which cannot be handled by traditional histogram-based algorithms:partial occlusions or pose changes, effective use of spatial information, tracking large targets own the same computational complexity as tracking small objects.Then, the paper analyzes two faults of robust fragments-based tracking and visual tracking via locality sensitive histogram:sensitivity of light change and fixed template is different to apply the various scenes. We introduce a locality sensitive histogram algorithm for visual tracking. Unlike the traditional density histogram by adding ones to the corresponding bin, a floating point will be added the corresponding bin in locality sensitive histogram for each position of an intensity value. It declines exponentially with the distance between the corresponding pixel location and the center of the histogram; thus the image of each pixel can is considered but those that are far away can be ignored due to the very small weights assigned. A robust visual tracking based on the locality sensitive histograms is composed by two main components:a feature for tracking that is robust to illumination changes and a multi-region tracking algorithm.Finally, fragments-based tracking and visual tracking via locality sensitive histograms are compared through nine groups of different characteristics data sets which are tested from two aspects:center location error(pixels) and tracking success rate. The final conclusion is that tracking based on locality sensitive histogram is more robust than fragments-based tracking, and adapt to more complex scenes.The research has achieved initial success, but there are still some problems which need to be solved. Such as target with fast moving, background clutter, long time occlusion and so on, they are still difficult to apply object tracking to practice. More efforts have to be done in the future.