Font Size: a A A

Moving Target Detection And Tracking In Complex Background

Posted on:2016-07-31Degree:DoctorType:Dissertation
Country:ChinaCandidate:J LiFull Text:PDF
GTID:1108330473461663Subject:Computer application technology
Abstract/Summary:PDF Full Text Request
Modern society develops rapidly, the probability of some harmful to the society events is in a rising trend. Simply rely to police force is difficult to curb the occurrence of those events. Thus the importance of intelligent video surveillance has become increasingly prominent. Intelligent video surveillance involves image processing, computer vision, pattern recognition and artificial intelligence and so on, which is a very challenging task.Intelligent video surveillance is very important in the area of image processing and computer vision research direction, and its core idea is to detect and track moving objects in complex background, and monitor the activities of objects in order to understand and describe the behavior of objects.This dissertation divides complex background into three kinds, according to the origin of problems, the range of influence and the length of duration,under which the moving object detection and tracking is analyzed. Then for solving moving object detection and tracking problems under complex background, such as light mutation, local disturbance, similar backgrounds and local occusion, the research of moving object detection and tracking is carried out.The main contribution and innovations of the dissertation includes:(1) Considering the influence of light mutation and local disturbance on moving target detection, a moving target detection algorithm based on CH feature(Contrast histogram)is designed. CH feature represents pixels, Using CH feature to represent pixels. And then it constructs background model based on CH feature by Gauss mixed model, initialises background model based on CH feature, trains model, and finally extracts the foreground.Because CH feature describes the symbiotic relationship between adjacent pixels, when grey value of the entire image block increases or decreases, the luminance information of CH feature remains unchanged, which reduces the light mutation and local disturbance influence.(2) Considering the influence of local disturbance on Gauss Mixture Model algorithm, a moving target detection algorithm based on Markov and Gauss Mixture Model is designed, which first constructs the Markov random walk model, uses BP algorithm to solve right value, and then obtain the edge feature.The edge feature is added to the traditional Gauss Mixture Model as the spatial information of image, and then the correlation of spatial information is further strengthened through morphological operation. The algorithm is robust under local disturbance.(3) Considering the influence of light mutation and the similar backgrounds on Mean Shift tracking algorithm, a Mean Shift tracking algorithm based on rotation invariant LBP feature is designed,which first calculates the rotation invariant LBP feature of target and candidate target, and then finds the position of Y in the plane, in order to obtain the minimum distance of target and candidate target in the rotation invariant LBP feature space, namely to obtain the maximum similarity. The new position of candidate target can be obtained through calculating the shift vector. Then the algorithm goes iteratively by using this new position as the initial position, which composes the tracking framework. This algorithm is robust to light mutation and similar backgrounds.(4) Considering the influence of local occlusion on particle filter tracking algorithm, an anti occlusion adaptive particle filtering tracking algorithm is designed.It adopts a rectangle as the tracking window, uses the K mean clustering algorithm to complete particle clustering in resampling, and then obtains the particle subgroup.It estimates the final state according to particles subgroups, and modifies the tracking window.When the areaSt(i)changes more than 5%, the tracking window maintains the same as the one in last frame. Otherwise, the tracking window will change according to the size of moving object, which is a self-adaptation process. At the same time it eases the degeneration problem of particle filter. This algorithm strengthens the robustness of tracking algorithm in case of local occlusion and moving object scale changing.
Keywords/Search Tags:Markov random walk, rotation invariant local binary pattern, anti- occlusion, KM clustering, Mean Shift, Particle Filter, moving object detection, moving object tracking
PDF Full Text Request
Related items