The Research Of Target Tracking Based On Monocular Vision

Along with the development of the theory of information processing and computertechnology, computer vision research has been more extensive and in-depth developed.Target recognition and tracking is one of the most important subjects in the field ofcomputer vision. It is a cross-disciplinary that integrating digital image processing,pattern recognition, and intelligent analysis. Therefor it has a wide range of applicationsin medical, military, navigation, traffic, mobile robot and other fields, particularlyplaying an irreplaceable role in terms of public safety and security.Currently, monitoring system in terms of the tracking moving targets, due to theuncertainty of the target movement, object occlusion as well as complex backgroundinterference, tracking algorithm is always can’t satisfy the requirements of accuracy,robustness and real-time.Therefore, this article has done in-depth research and improvedsome tracking algorithm based on OpenCV open source vision library and traditionalalgorithm to achieve goals.First, in terms of image pre-processing for noise problem, put forward a fastadaptive switching median filter algorithm, dividing the image in accordance with thedifferent degrees of the noise density, so that to use different processing window.Apartfrom this, group the pixels firstly to speed up the processing speed when caculating themedian value. As for moving target detection, on the basis of adaptive Gaussian mixturebackground modeling algorithm, an improvement is obtained according to the numberof moving target adjusting the Gaussian model number and use different learning ratefor background and moving region. The existence of the shadow leads to the change ofthe shape of the objects, so a shadow suppression algorithm is achieved, improving thedetection accuracy and robustness ultimately.As for single target tracking, the thesis proposes a target tracking method based onthe Kalman filter and mean shift to solve occlusion problem. When there is an occlusion,using Kalman filter predicting target location. For multi-target tracking, combine theblob geometric characteristics with color moment as multi-feature fusion to track target.Gaussian mixture modeling algorithm to obtain a plurality of blob of target, and then analyze the geometric characteristics of these individual blob, using Kalman filtering topredict, due to the similarity between the objects, the prediction result is not entirelyaccurate, thus combining the image color moments correcting the judgment, theexperimental results show that the proposed multi feature fusion target trackingalgorithm has good robustness.As for train velocimetry system, when using traditional Hough line to detect traintrack, add the conditions of the the track color contrast judgment as well as otherrestrictions to improve the result. By setting the region of interest and virtual triggercoils, ultimately measure the speed of the train.