车辆主动安全中关于车辆检测与跟踪算法的若干研究

发布时间:2018-06-20 02:53

  本文选题:智能车 + 车辆检测 ; 参考:《吉林大学》2015年博士论文


【摘要】:本文围绕着车辆主动安全中的车辆检测和车辆跟踪技术中面临的一些关键问题展开了深入的研究,并取得了一定的进展,具体包括: 在基于车载相机的车辆检测方面,通过对采集图像中车辆的特征进行分析,提出了一种基于水平边缘车辆波的车辆检测方法,该方法中的水平边缘车辆波不但能很好地描述车辆的特征,同时也能排除大量干扰情况。方法的准确率和处理效率都很高,且能同时应用于前向和后向车辆检测。另外,根据图像中车辆尾灯的特点,,本文提出了一种车辆尾灯的车辆检测方法,方法可作为多特征车辆检测的一个判据。针对单一特征误检率高的问题,本文提出了两种多特征融合的车辆检测方法,一种是基于投票法的多特征融合车辆检测方法,方法能极大程度地降低误检率。另一种是基于运动轨迹的多特征融合车辆检测方法,方法使用轨迹判断代替分类器判断,解决了车辆检测分类器依赖于样本规模的问题。为了达到提高检测率的同时降低误检率的要求,本文提出了两种基于决策理论的多特征融合车辆检测方法。首先将三个基于特征的车辆检测方法(对称性、车尾灯,HoG+AdaBoost分类器)的输出结果进行模糊化表达,之后分别使用Choquet模糊积分和D-S证据理论对模糊化的车辆特征算法结果进行结果融合。通过特征融合,实现了多工况下的车辆鲁棒性检测,提升了基于车载视觉传感器的车辆检测算法的性能。 在多车辆跟踪方面,本文重点解决了尺度不断变化,目标消失以及跟踪过程中目标部分或全部遮挡情况下的多车辆跟踪问题。首先针对尺度不断变化的多车辆跟踪问题,本文提出了一种基于扩展卡尔曼滤波以及基于On-line boosting的离线和在线学习相结合的跟踪方法,在对车辆进行跟踪的过程中结合离线多尺度车辆表观分类器及遮挡判断机制来处理目标消失以及目标间的严重遮挡问题,试验结果表明,本文提出的算法在保证实时性的同时实现了尺寸不断变化条件下的车辆跟踪。除此之外,为解决现有的基于粒子滤波的目标跟踪方法不能满足复杂工况下多车辆跟踪的问题,本文提出了一种基于改进粒子滤波的车辆跟踪方法。方法中的适应于移动平台下的多车辆跟踪问题的粒子滤波状态方程,基于归一化MCRP面积的目标初始和消失处理方法以及跟踪过程中的目标位置冲突处理方法保证了复杂工况下多车辆跟踪方法的鲁棒性。 通过在公开测试集和自行制作的测试集上对本文提出的基于车载视觉传感器的车辆检测和车辆跟踪方法进行测试,并与目前一些广泛使用的方法进行对比,验证了本文算法的有效性。
[Abstract]:In this paper, some key problems in vehicle detection and vehicle tracking technology in active vehicle safety are studied, and some progress has been made. In the aspect of vehicle detection based on vehicle camera, a vehicle detection method based on horizontal edge vehicle wave is proposed by analyzing the characteristics of the vehicle in the collected image. The vehicle waves at the horizontal edge in this method can not only describe the vehicle characteristics well, but also eliminate a large number of disturbances. The method has high accuracy and processing efficiency, and can be applied to both forward and backward vehicle detection. In addition, according to the characteristics of vehicle taillights in the image, this paper presents a vehicle detection method for vehicle taillights, which can be used as a criterion for multi-feature vehicle detection. Aiming at the problem of high false detection rate of single feature, this paper proposes two vehicle detection methods based on multi-feature fusion, one is multi-feature fusion vehicle detection method based on voting method, which can greatly reduce the false detection rate. The other is a multi-feature fusion vehicle detection method based on motion trajectory, which uses trajectory judgment instead of classifier to solve the problem that vehicle detection classifier depends on sample size. In order to improve the detection rate and reduce the false detection rate, two multi-feature fusion vehicle detection methods based on decision theory are proposed in this paper. Firstly, the output results of three feature-based vehicle detection methods (symmetry, tail light HoG AdaBoost classifier) are expressed by fuzzy method. Then Choquet fuzzy integral and D-S evidence theory are used to fuse the result of fuzzy vehicle feature algorithm. By means of feature fusion, vehicle robustness detection under multi-working conditions is realized, and the performance of vehicle detection algorithm based on vehicle vision sensor is improved. In the aspect of multi-vehicle tracking, the problem of multi-vehicle tracking with varying scales, vanishing targets and partially or totally occluded targets in the tracking process is solved in this paper. First of all, aiming at the problem of multi-vehicle tracking with changing scales, this paper proposes a tracking method based on extended Kalman filter and On-line boosting, which combines offline and online learning. In the process of vehicle tracking, the problem of object vanishing and serious occlusion between targets is dealt with by combining off-line multi-scale vehicle surface classifier and occlusion judgment mechanism. The experimental results show that, The algorithm proposed in this paper not only guarantees the real-time performance, but also realizes the vehicle tracking under the condition of constant size change. In addition, in order to solve the problem that the existing particle filter based target tracking method can not meet the problem of multi-vehicle tracking under complex operating conditions, a vehicle tracking method based on improved particle filter is proposed in this paper. The particle filter equation of state for multi-vehicle tracking on mobile platform is proposed in this paper. The target initial and vanishing methods based on normalized MCRP area and the target position conflict processing method in the tracking process ensure the robustness of the multi-vehicle tracking method under complex operating conditions. The vehicle detection and tracking methods proposed in this paper are tested on the open test set and the self-made test set, and compared with some widely used methods. The validity of the proposed algorithm is verified.
【学位授予单位】:吉林大学
【学位级别】:博士
【学位授予年份】:2015
【分类号】:TP391.41

【参考文献】

相关期刊论文 前6条

1 齐美彬;潘燕;张银霞;;基于车底阴影的前方运动车辆检测[J];电子测量与仪器学报;2012年01期

2 丁莹;李文辉;范静涛;杨华民;;基于Choquet模糊积分的运动目标检测算法[J];电子学报;2010年02期

3 刘晨光;程丹松;刘家锋;黄剑华;唐降龙;;一种基于交互式粒子滤波器的视频中多目标跟踪算法[J];电子学报;2011年02期

4 王莹;李文辉;;基于多特征融合的高精度视频火焰检测算法[J];吉林大学学报(工学版);2010年03期

5 李文辉;周强;王莹;张德才;;基于均值偏移粒子滤波的自适应跟踪[J];吉林大学学报(工学版);2012年02期

6 解文华;肖进胜;易本顺;张亚琪;李明;;一种基于Mean Shift和C-V模型的车辆跟踪算法[J];湖南大学学报(自然科学版);2012年07期

相关博士学位论文 前1条

1 王楠;基于多视觉特征融合的后方车辆检测技术研究[D];东北大学 ;2009年



本文编号:2042610

资料下载
论文发表

本文链接:https://www.wllwen.com/shoufeilunwen/xxkjbs/2042610.html


Copyright(c)文论论文网All Rights Reserved | 网站地图 |

版权申明:资料由用户4405e***提供,本站仅收录摘要或目录,作者需要删除请E-mail邮箱bigeng88@qq.com