Research On Improved Object Tracking Algorithm In Complex Enviornment

With the rapid development of China’s transportation, people’s lives become morecomfortable and convenient. But at the same time, road congestion, air pollution, andtraffic accidents become increasingly serious problems. The planning and construction ofthe intelligent transport system is the key measure to solve traffic problems all over theworld. And object tracking is one of the key technology to realize intelligent transportsystem. Although object tracking methods have been used exhaustively in manyresearchers’ work, given rise to multiple applications, some difficulties of object trackingare not completely solved, such as occlusion and object model updating. Object detectionand tracking algorithm under complex environment is studied theoretically andexperimentally in this thesis, which is based on advanced traffic management system.Firstly,an object detecting method based on adaptive Gaussian mixture model isproposed, which eliminates object’s shadow by transforming RGB color space into HSVand reduces the influence of illumination condition changes by updating coefficientadaptively. Convincing results have been obtained in sequences under illuminationcondition changes or minor disturbances in background. The method has a strongadaptability to complex scenes, and is suitable for video monitoring system both indoorand outdoor.Secondly, a method that bandwidth window in tracking process changes adaptivelyaccording to object scale is presented. In the traditional Mean Shift algorithm, the size ofbandwidth window remains unchanged through the whole tracking process, which cannotchange with the scale of the object and always results in poor tracking. This thesisenhanced the Mean Shift tracker to regard contour size of the object as bandwidth windowsize, and automatically update the window size according to the contour.Thirdly, an algorithm which combining the improved Mean Shift algorithm andKalman filter is proposed to solve occlusion problem. The Mean Shift tracker is a widelyused tool for robustly and quickly tracking, but the state obtained from Mean Shiftalgorithm may not be the local extreme state under occlusion, such as target region is partially occluded or completely occluded, which usually results in target loss. Kalmanfilter is successfully used to predict the object position under occlusion. Firstly, accordingto the target location in the previous previousprevious frame, Kalman filter predicts targetlocation in the current frame adaptively.Secondly, find the real target location in theneighbourhood by mean shift algorithm.Finally,update the filter parameters. Because theadaptive Kalman filter predicts target location through system equation, it can improve thetracking effect in occlusion in a certain degree.Then, Mean Shift algorithm based on mixed histogram is studied in complex scenes.Color feature is invariant to scaling, translation, rotation and other conditions. However, itproduces, in many cases, a poor representation of the target, which might result in poortracking. it is often unstable and inadequate in complex scenes, such as illuminationcondition change, target posture change, target color change, and so on. To enhance thetracking, we introduce a method to use multiple reference histograms for producing amixed histogram which is more appropriate for object tracking. Regarding referencehistogram convex as the target model, the tracker finds target region by narrowing thedistance between color histogram and convex reference histogram.Finaly, an adaptive Kalman filter with model error and noise error is proposed topredict the target position. In ideal condition, Kalman filter can obtain linear unbiasedminimum variance estimation. However, in practical condition, it often cause filterdivergence in some conditions such as lack of prior knowledge, mathematical model doesnot comply with practical condition, neglect some coupling component, system noise andmeasurement noise is improperly selected, and so on. Aiming at solving the mentionedproblem, an adaptive Kalman filter with model error and noise error is studied in thisthesis. The adaptive Kalman filter calibrates model parameters, noise statistics and gainmatrix through online estimation by using observation data, which aims at reducing theestimation error and improve precision of the filter.