基于图像运动区域的车辆遮挡跟踪算法

A Region-based Vehicle Tracking Algorithm under Occlusion

针对区域跟踪算法难以解决因车辆遮挡而引起误检的问题,提出了基于图像运动区域的车辆跟踪算法采用背景剪除法提取运动区域,通过计算相邻帧运动区域的位置变化实现区域跟踪;建立车辆的二维矩形框模型,分析“区域——车辆”关系,结合区域跟踪的结果来判定车辆之间是否发生遮挡,并根据车辆行为来初始化车辆模型轮廓及速度;采用Kalman滤波器预测车辆在当前帧的位置,并以此预测位置作为车辆模型的初始位置进行模型轮廓的自适应调整,得到模型新的矩形轮廓;将新轮廓其所确定的几何中心位置作为测量值反馈回Kalman滤波器,修正Kalman系数,进行自回归运算和计算最佳匹配位置,从而实现车辆跟踪.算法测试实验使用的视频采集自江苏省通启高速公路视频监控系统,采用P4/2.4单CPU,结果表明,在为25帧/s视频流下,该算法准确跟踪率达到94.72%,有效解决遮挡问题,并具有较好的鲁棒性.

A novel algorithm is presented for real-time vehicle tracking in image sequences acquired by a stationary camera, to lower the error rate under occlusion by the previous region-based tracking methods. Vehicle tracking has already been a hotspot in the field of computer vision and image recognition, and it is the foundation of the traffic control system. Motion regions are obtained by thresholding the result of subtracting the image from the background. The variety of the regions in the two adjacent images can be defined as four particular region behaviors which are calculated from their locations in the frames. When this relationship has been established, we use 2- dimensional rectangle patch to model the car, which adequately resembles the car＇s shape and dynamics, and analyze the region-car relation through region behaviors. Considering different car-region relations, we use defined regulations to judge whether the car occlusion has occurred, and then initialize the velocity and contour of the car model according to different cases. The initial location of the car model is given by the Kalman filtering prediction. Based upon the original parameters of the model, both geometrical center position and contour, active contour adjustment is taken to gain a new contour which is fitter the real size. The adjustment is to move the edges of the rectangle patch model to satisfy the defined constraint, making the counter fit to the car. This new contour will determine the center of the rectangle patch, which will be feedback to the Kalman filtering as the measurement to modify the Kalman gain, estimate the optimum position, and continue the estimation in the next frame. And this optimum position is the final outcome of the algorithm in the current frame and thus it is the vehicle tracking result. The testing experiment was implemented on the computer equipped with P4/2. 4M Hz single CPU, by a video sequence recorded in the Tongqi express way （Jiangsu province） monitor system. The correct tracking rate was 94. 72% of 25 frames per second, and the error rate was dynamically decreased under partial or heavy occlusion compared with previous region-based methods. Experimental results based on variety scenes illustrate that our algorithm is effective and robust.