Researches On Moving Vehicle Detection And Tracking In Video Based On Bachground Subtraction Approach And Its Software Design

Traffic detection and measurement is one of the basic tasks in intelligenttransportation systems. Among various traffic detection problems, the accuratedetection of moving targets in video clips is one of the most important ones. It is also aresearch hot spot in the field of computer vision as well as in intelligent transportationsystems.A number of motion detection methods have been proposed. The most typicalones are the light flow methods, the frame difference methods, and the backgroundsubtraction methods. Gaussian mixture background model based on the backgroundsubtraction is among the most frequently adopted detection algorithms. Gaussianmixture background model is the combination of several single Gaussian models.However, one single Gaussian model is not enough to cover the changes of thebackground, and is not desirable to describe the disturbances in the background.Though adaptive to slow changes of the global illumination in the view field, themethod is vulnerable to fast changes of the illumination. The thesis focuses on theimpacts of sudden illumination changes, and an improved Gaussian mixturebackground model algorithm is studied.Improper updating rate in the Gaussian mixture model methods can lead to thedetection of false foreground objects due to the sudden changes in illumination. This isbecause the model cannot follow the change of the background, thus the significantchanges over most of the pixels in the video frame is likely to be mistaken forforeground. If the scene is interpreted from the perspective of the HSV color model, itcan be reasonably assumed that the global illumination change affects only the Vcomponent, while leaving the H and S components unchanged. From this assumption,the proposed method uses the average change of V component, VPa, as an indicator forsudden global illumination change. According to the value of VPa, the current framecan be categorized into one of the following three situations: minor illuminationchange, considerable illumination change, and sudden illumination change. For minorillumination changes, the background model is updated in the standard way. Underconsiderable illumination changes, the background model is updated using a greaterupdating rate. If sudden illumination change is the situation, the V component over thewhole frame is adjusted accordingly, and the modified V component along with the unchanged H and S components are transformed into the RGB color model, and thebackground model is updated using the standard process. By this treatment, thesignificant changes in the RGB values caused by the sudden illumination change canbe avoided, resulting in a smoother updating of the background. Experimental resultsshow the effectiveness of the proposed method.Tracking method based on the Kalman filter and target matching algorithm basedon Mean-shift is introduced and realized through a combination of the two methods.Software for vehicle detection and tracking in videos is designed and implementedusing Visual C++2008with OpenCV image processing library. The software can notonly demonstrate the original video and the results of video detecting and tracking andprovide a number of detection and tracking algorithm, including improved backgroundminus division paper proposed. In addition, user can set criterion motion trail of ROIof vehicle through the interactive operation in software for evaluations andcomparisons of experimental results.