Research And Implementation On Algorithm Of Accurate Contour Tracking Based On Video

Target tracking is one of the essential tasks in computer vision research.There are many computer vision applications based on the target tracking technology,such as intelligent monitor,action recognition,augmented reality,and so on.This thesis proposed a new method to tracking the contour of the target which used sparse low rank constraint to measure the distance between super pixels,the following is the main contant of this thesis.(1)Research on the method of measuring distance between super pixels based on the sparse low rank constraint.According to the continuity of the continus frames in the video and the continuity of the super pixels in one frame,we build a linear model,one super pixel in the current frame can be represented as a linear conbination of the previous frames in the video,and the linear coefficient should be sparse and low rank.linear coefficient can measure the simarity between super pixels very well.(2)We propose an argorithm to tracking the target's contour which based on the super pixel simarity.By computing the significant target probablity by using super pixel simarity,and combining with the appearance target probablity based on the distribution of YUV color space,we can constract a condition random field model to segment the tracking target,and release the terget's contour tracking.We analysis the good ability of this algorithm to handle the occlusion and scale case.(3)We discuss the effect of extracting feature from the deep convolution neura network when the feature is used in this algorithm.While using the VGG-16 network to extract the feature from the video,because the early layers has higher spatial resolution and later layers has more semantics,so we use the output of the first big convolution of the VGG-16 network as the feature of this algorithm,and get better tracking result than using the mannual designed feature.(4)Extend the single target contour tracking model to the muti target contour tracking model.Because the model itself can be naturally extended from single target to muti target,so it will not cost too much more compute resource to tracking muti target than single target.