Research On Visual Tracking And Moving Object Detection Based On Optical Flow Feature

Moving object detection (MOD) and Visual Object tracking (VOT) is one of thechallenging and typical research domain of Computer Vision (CV), it play an significantrole in the application of guidance weapon, aerial photo and visual supervision. This paperfocuses on MOD and VOT area based on the research of predecessor in the CV domain. Ihave three contributions about CV.(1) This paper proposed a moving object detection algorithm based on optical flow[1].First, this method is improved on the base of LK optical flow (Lucas Kanade), then, thedetection algorithm combines the LK and classical Mean Shift together to estimate thebackground moving vector, this estimation can effectively detect the moving object in thedynamic background, and overcome the drawbacks of cavity effect of frame subtractionand background compensation. Second, this method abandon the uniform samplingmethod in the reason of flow shift in the pixels which is lack of text feature. I propose aninhomogeneous sample method based on local variance and resample method thatimprove the successful detection rate, optical flow usage rate and algorithm speed of smallmoving object. Third, to compensate the side effect of LK algorithm that is difficult todetect the slowly moving object, I used the saliency detection algorithm to assist the LKmethod which can detect the slowly moving objects successfully.(2) This paper proposed a object tracking method based on local text feature andglobal color feature. This method is improved on the base of Compressive SensingTracking[2]that we have three improvement than the origin method. First, we use theanother Haar-like feature to describe the target, which can describe to the globalinformation. Second, the origin object features is sensitive to the local text image, whenthe target is transforming or rotating, it may be lose the target. We implement the methodby adding another dim of color feature that make the object features more balance betweenglobal feature and local feature. The Haar feature can capture the small changes of objectso that it can recognize the partial shelter of the object, the color feature we propose canadapt to the rotation and zoom transform, and make the final result much better. Third, weuse the particle filter to search for the object instead of the ergodic search in the originmethod, so lots of the redundant computation is cut and the speed of algorithm becomesfaster.(3) This paper proposed a fast algorithm to calculate the local feature of image, whichcan calculate the local sum, local square sum, variance, standard deviation, entropy andmean-value filter of image pixels in O(1) time complexity, and can calculate the cosinedistance and correlation between two image sub-region in O(n) time complexity.