车载视频监控中基于乘客检测和跟踪的客流计数方法

发布时间：2018-09-17 18:18

【摘要】：客流计数是车载智能视频监控系统运作的基础,通过客流计数,可以实时地获取各路公交车在各个时段的上下车人数,方便维护公交车秩序和安全。然而,由于公交车上的复杂环境,车体的抖动,图像的失真,乘客间的相互遮挡等问题,使得目前还没有一套实时的、精确的客流计数方案。针对以上问题,本文提出一种基于图像处理和乘客检测及跟踪技术的客流计数方法。首先对视频进行稳像,减小公交车抖动引起的图像序列间的偏移,然后对图像进行梯形校正,减小摄像头角度引起的图像梯形失真,最后对乘客进行检测和跟踪,解决公交车上乘客相互遮挡和光照变化明显的问题,并根据乘客的运动方向判断乘客上下车情况,实现客流计数。首先,在当前抖动图像的左上角和右上角选取两个子块,进行局部运动估计,并用局部运动矢量的平均值作为运动补偿矢量,减小视频抖动。其次,根据摄像头倾斜角度,计算得到透射变换矩阵,对失真图像进行视角变换,并通过双线性插值算法对图像进行梯形校正。然后,对待检测图像进行自适应阈值背景差分,实现乘客目标分割,去除背景中类似人体轮廓物体的干扰;提取乘客头肩部梯度方向直方图(Histogram of Oriented Gradient,HOG)特征,训练支持向量机(Support Vector Machine,SVM)头肩分类器,实现乘客目标检测;对乘客进行基于快速鲁棒性特征(Speeded-Up Robust Feature,SURF)的目标跟踪。最后,对视频图像设置计数线,判断乘客是否跨越计数线,并根据乘客运动的方向,统计出各个站点的上下车人数和车厢乘客总人数。
[Abstract]:The passenger flow count is the basis of the intelligent video surveillance system. Through the passenger flow count, the number of buses can be obtained in real time, and it is convenient to maintain the order and safety of the bus. However, due to the complex environment on the bus, the jitter of the body, the distortion of the image and the mutual occlusion among the passengers, there is still no real-time and accurate passenger flow counting scheme. In view of the above problems, this paper presents a passenger flow counting method based on image processing and passenger detection and tracking techniques. Firstly, the video is stabilized to reduce the deviation between image sequences caused by bus jitter, and then trapezoidal correction is carried out to reduce the trapezoidal distortion caused by camera angle. Finally, passengers are detected and tracked. The problem of mutual occlusion and obvious change of illumination on bus is solved, and passenger flow count is realized by judging the situation of passengers getting on and off according to the movement direction of passengers. Firstly, two sub-blocks are selected in the upper left corner and the upper right corner of the current jitter image to estimate the local motion, and the average value of the local motion vector is used as the motion compensation vector to reduce the video jitter. Secondly, according to the tilt angle of the camera, the transmission transformation matrix is obtained, and the distorted image is transformed into the angle of view, and the trapezoidal correction of the image is carried out by bilinear interpolation algorithm. Then, the detection image is treated with adaptive threshold background difference to achieve passenger target segmentation, remove the interference similar to human contour objects in the background, and extract the (Histogram of Oriented Gradient,HOG features of passenger head and shoulder gradient direction histogram. Support vector machine (Support Vector Machine,SVM) head-shoulder classifier is trained to realize passenger target detection and passenger target tracking based on fast robust feature (Speeded-Up Robust Feature,SURF) is carried out. Finally, the counting line is set to determine whether the passengers cross the counting line, and according to the direction of passenger movement, the number of passengers in and out of each station and the total number of passengers are calculated.
【学位授予单位】：天津大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：U463.67;TN948.6

【参考文献】