复杂环境下基于视频的车辆跟踪算法研究

发布时间：2018-03-06 11:47

本文选题：车辆跟踪　切入点：车辆检测　出处：《东南大学》2015年博士论文　论文类型：学位论文

【摘要】：交通信息采集是交通管理、控制以及诱导的基础。相比传统环形线圈检测技术,基于视频的车辆跟踪技术具有安装维护方便、检测区域大、可检测参数多等优点,在交通信息采集方面具有巨大的潜力,但是车辆跟踪的性能受复杂环境的影响很大,比如交通场景中光照的变化、车辆间的遮挡现象等因素都会影响车辆跟踪的可靠性与准确性。因此,复杂环境下基于视频的车辆跟踪技术研究对于交通信息采集具有重要意义。本文对环境中存在光照变化和车辆间遮挡现象时的车辆跟踪算法进行研究,主要研究内容如下：(1)为了在光照变化的情况下检测运动车辆,提出了自适应ViBe(visual background extractor)背景模型。分析了ViBe背景模型存在的问题：在光照变化情况下,当阈值过小时部分背景区域会被误检测为车辆,而当阈值过大时部分车辆区域会被误检测为背景。为解决这一问题,定义了车辆检测误差函数。以被误检测为车辆的背景区域面积与图像面积之比作为第一类误差函数,用以衡量阈值过小的影响；利用包含车辆区域的最小凸包近似地表示真实的车辆区域,以该凸包为参考,计算图像中所有车辆的相对误差的平均值并以之作为第二类误差函数,用以衡量阈值过大的影响。最后,在首先减少第二类误差、其次减少第一类误差的准则下,设定一组条件以判断阈值的合理性,并利用步长法对其进行调整。实验结果表明,自适应ViBe背景模型能够提高光照变化情况下车辆检测的准确性。(2)为了提高车辆跟踪的持久性和准确性,提出了基于非网格块运动矢量的车辆跟踪算法。分析基于网格块运动矢量车辆跟踪算法的缺点在于：一、在网格分块下,较小的车辆区域中包含的块的数量过少,导致噪声运动矢量的影响增大、车辆跟踪的持久性下降；二、在块的扩展环节,目标车辆容易受噪声区域和附近车辆的影响,导致车辆跟踪的准确性下降。针对上述问题,一方面在分块方式上提出以非网格的方式对车辆进行分块,从而使得块可以处于图像中的任何位置,并且块与块可以重叠。这样能够使较小的车辆区域也包含足够多的块,从而减少噪声运动矢量的影响,提高车辆跟踪的持久性。另一方面在块的移位环节对每个块分别根据其自身运动矢量对其进行移位。为此,根据块运动矢量在空间分布上的延续性,提出了基于邻域的噪声运动矢量检测和调整方法。这种移位方式避免了对块进行扩展,在移位中减少了噪声区域和附近车辆对目标车辆的影响,从而提高了车辆跟踪的准确性。实验结果表明,相比基于网格块运动矢量的车辆跟踪算法,基于非网格块运动矢量的车辆跟踪算法具有更好的跟踪持久性和准确性。(3)提出了非网格分块下的Markov随机场遮挡处理算法。当遮挡车辆和被遮挡车辆的运动矢量相接近时,网格分块下的Markov随机场遮挡处理算法难以对遮挡区域进行准确分割。针对这一问题,在基于非网格块运动矢量的车辆跟踪算法的框架下,利用欧式距离定义了非规则的时空邻域系统,以描述非网格分块中块与块之间的几何关系,并以非规则的时空邻域系统为基础,建立了非网格分块下的Markov随机场模型。在能量函数定义方面,构建了由运动灰度差与颜色直方图距离组成的二维高斯向量,以融合车辆的运动信息与颜色信息,更为完整地表达遮挡车辆与被遮挡车辆之间的差异；最后利用模拟退火法对能量函数进行优化,以实现遮挡分割。实验结果表明,非网格分块下的Markov随机场遮挡处理算法在两辆车的运动矢量相接近时能够利用颜色信息的差异处理车辆遮挡现象,扩大了车辆遮挡处理的范围,提高了车辆遮挡处理的准确性。(4)针对车辆刚进入交通场景就发生遮挡的问题,提出了基于长轴方向固定椭圆拟合法的遮挡初始分割算法。首先,通过前景凸包比来检测车辆遮挡现象,并利用一组与车道线方向平行的分割直线将前景轮廓分割为若干组子轮廓。之后,在最小二乘椭圆拟合法的基础上增加了长轴方向固定这一约束,提出了长轴方向固定的最小二乘椭圆拟合法,利用它对每一组子轮廓进行椭圆拟合。最后,利用牛顿迭代法定义了椭圆拟合的评价函数,并通过最小化该评价函数以寻找最优分割直线,从而获得车辆的初始位置。实验结果表明,该算法在遮挡初始分割中对于不规则的前景轮廓具有较好的鲁棒性,能够得到准确的车辆初始位置。(5)为了验证本文算法的有效性,在Matlab2011环境中设计并实现了车辆跟踪软件。该软件能够对交通场景中的车辆进行检测和跟踪,并利用车辆的轨迹信息检测交通量以及停车、变道等交通行为。
[Abstract]:Traffic information collection is traffic management, control and induction basis. Compared with the traditional loop detection technology, video vehicle tracking technology has the advantages of convenient installation and maintenance based on the detection area, can detect many parameters and other advantages, has great potential in traffic information collection, but the performance of vehicle tracking is affected by the complex environment, such as traffic scene illumination changes, the factors of vehicle occlusion phenomenon will affect the reliability and accuracy of vehicle tracking. Therefore, the complex environment of video vehicle tracking technology research is of great significance for traffic information collection. Based on the illumination change and inter vehicle occlusion when the vehicle tracking algorithm is studied in this paper to exist in the environment, the main research contents are as follows: (1) in the case of illumination variations of vehicle detection, the adaptive ViBe (visual backgroun D extractor) background model. Analyzed the existing ViBe background model: under illumination changes, when the threshold is too small will be part of background error detection for the vehicle, and when the threshold is too large when the vehicle area will be mistakenly detected as the background. In order to solve this problem, the vehicle detection error function is defined to be the background. Error detection area and image area ratio of the vehicle as the first class error function is used to measure the threshold is too small; with minimum convex hull of the vehicle region similar to that of the real vehicle area, with the convex hull as the reference, the relative error calculation of all vehicles in the image of the average value and as the second kind of error function to use, in order to measure the impact threshold is too large. Finally, in the first second categories of reduce error, then reduce the first error criterion, setting a set of conditions to determine the reasonable threshold , and to adjust it by step method. The experimental results show that the adaptive ViBe background model can improve the accuracy of vehicle detection under the condition of illumination change. (2) in order to improve the durability and accuracy of vehicle tracking, tracking algorithm is proposed for non grid block motion vector based vehicle tracking algorithm. Analysis of grid block motion vector the vehicle based on the disadvantage is: first, in the grid block, block the small number of the vehicle contains smaller region, resulting in the increase of the noise impact motion vector, persistent tracking vehicles dropped; two, in the extended link block, the target vehicle is easily affected by the noise and the area near the vehicle, resulting in a decline in accuracy vehicle tracking. To solve these problems, on the one hand, the block mode of vehicle block to non grid, so that the block can be in any position in the image, and The blocks can be overlapped. This can make the vehicle the small area also contains enough block, thereby reducing the influence of noise of the motion vector, improve persistent tracking vehicles. On the other hand, shift link piece for each block respectively according to the shift of its own motion vector. Therefore, according to the continuity of the movement the vector in the spatial distribution of forward motion vector noise detection and adjustment method based on the neighborhood. This way to avoid the shift block expansion, reduce the influence of noise in a shift in the region and near the target vehicle from the vehicle, and improve the accuracy of vehicle tracking. The experimental results show that compared with the tracking algorithm of grid blocks the vehicle motion vector based on non grid block motion vector of the vehicle tracking algorithm has better tracking accuracy and durability. (3) based on the proposed non grid Markov block with the airport Occlusion handling algorithm. When the motion vector blocked vehicle and vehicle occlusion are close, mesh Markov under random occlusion processing algorithm is difficult to accurately occlusion segmentation. Aiming at this problem, based on the frame tracking algorithm for non grid block motion vector of the vehicle, using the Euclidean distance definition of space-time the neighborhood system of irregular, non grid points to describe the geometric relationship between the block and the block, and the spatio-temporal neighborhood of irregular basis to establish the non grid Markov block under random field model. The energy function definition, construct the two-dimensional Gauss vector composed of gray and color difference movement the histogram distance, the motion information and color information fusion of the vehicle, a more complete expression of occlusion vehicles and blocked differences between vehicles; by using the simulation to optimize the energy function of the annealing method, In order to achieve occlusion segmentation. The experimental results show that the difference of non grid Markov under random occlusion is close to the motion vector processing algorithm for two cars to use the color information of the vehicle occlusion phenomenon, expand the scope of vehicle occlusion treatment, improve the accuracy of the vehicle cover block processing. (4) for vehicles just enter the traffic scene occlusion problem, put forward the direction of the long axis of the ellipse fitting fixed initial segmentation algorithm based on block method. First, the prospect to detect vehicle occlusion phenomenon of convex hull, and using a group and lane to the parallel line segmentation the foreground contour segmentation into several sub group profile. After at least square ellipse fitting based on the increase in the long axis direction of the fixed constraint, put forward the long axis of least square ellipse fixed fitting, using it to each sub contour in ellipse Fitting. Finally, the evaluation of ellipse fitting function is defined using the Newton iterative method, and by minimizing the evaluation function to find the optimal partition line, so as to obtain the initial position of the vehicle. The experimental results show that the algorithm in the initial block segmentation for irregular foreground contour has good robustness, can get the initial position of the vehicle accurate. (5) in order to verify the effectiveness of the proposed algorithm, the design and implementation of vehicle tracking software in the Matlab2011 environment. The software is capable of detecting and tracking in traffic scene and the use of vehicles, traffic volume and vehicle parking detection track information, change lanes and other traffic behavior.

【学位授予单位】：东南大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：U495

【相似文献】