Infrared Vehicle Tracking Algorithm Research Based On Svm And Mean-Shift

Vehicle detection and tracking is an important and challenge issue in Intelligent Transportation Systems (ITS) and Driver Assistance System (DAS) and it has been broadly investigated in the past years. Determining the position of other vehicles on the high speed road is important to help Driver Assistance System. Thus, robust and reliable vehicle detection and tracking are necessary in these systems. Recently, active researches on vehicle detection and tracking are done for Driver Assistance System-a collision warning and avoidance, vision enhancement, etc. The vehicle detection and tracking algorithm for DAS require a robust feature extraction and tracking method regardless of light and road condition and an exact estimation of vehicle position and velocity regardless of the distance from the ego-vehicle. However, most research was carried out in visual video which was sensitive to light and had serious ongoing vehicle shadow problems, thus it is difficult to extract the accurate vehicle contour when detecting. In this paper we focus on vehicle detecting and tracking in infrared videos. Compared with normal visual videos, infrared videos are robust to day/night changes and hence suitable for environments with poor or unstable lighting conditions.In this paper, vehicle tracking under static background are researched by SVM and Mean-shift. Four main research works have been done as follows:(1) Infrared vehicle target enhancement based differential evolution algorithm and stationary wavelet transform. Considering noise and low contrast of infrared image, an efficient nonlinear adaptive enhancement algorithm, which is based on differential evolution algorithm and stationary wavelet transformation, is proposed. Evaluation function is constructed by combing information entropy, signal-noise-ratio with standard deviation of enhanced image. A nonlinear transformation function is designed to enhance the contrast of the infrared image according to the properties of the infrared image. The optimal nonlinear transformation parameters are determined by combining differential evolution algorithm with the constructed evaluation function. The infrared vehicle image enhancement examples illustrated that the proposed algorithm is better than multi-scale nonlinear enhancement algorithm, stationary wavelet nonlinear enhancement algorithm and histogram equalization algorithm in overall performance.(2) Infrared vehicle detection algorithm based on Grow-Cut and least squared fitting. Firstly, Grow-Cut and Least Squared Fitting were used to detect lanes. Some tests were done using infrared vehicle image to detect lanes, which show that it accurate lane with no mistake lane are got by selecting seed points to segment ROI. For low resolution image, it could get good result too. Secondly, highlight region in the ROI are found by threshold; vehicle statistic length and width are used to conform rough vehicle position region box which see highlight region as center; in the rough vehicle position region, edge detection are used to further determine position of the vehicle, then accurate vehicle position are got. Some experiments proved that the method is viable and compared with traditional vehicle detection method, we could get more complete vehicle.(3) Infrared vehicle classification and tracking based on SVM and mean-shift. Firstly, extract features of the detected infrared vehicle image, divide images and their features into two parts separately as training sample and testing sample, use SVM to train the training sample and classify the testing samples to test. Experiment shows that error rate using SVM is less than that of using BP neural network classifier and RBF neural network classifier. Thus Kalman filter and Mean-shift are combined to trace the vehicle target in the infrared image. The overall performance of the proposed algorithm is better than Kalman filter and Mean-shift method.