交通视频中车辆多目标跟踪与特征提取的研究

发布时间：2018-03-18 09:05

本文选题：运动车辆检测　切入点：改进的ViBe算法　出处：《天津工业大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着人工智能的发展,在智能交通系统中,计算机视觉技术已经融入到智能交通视频的分析中。但由于交通视频中背景复杂,噪声较多,车辆运动目标不规律,使得运动目标的检测和车辆的多目标跟踪仍然面临着诸多具有挑战性的问题。同时在越来越多的车辆视频信息中,如何对检测跟踪到的车辆做有效的特征提取和方便快捷的检索比对。这都成了智能交通信息领域迫切需要解决的问题。本文对交通视频中运动目标检测、车辆多目标跟踪以及目标特征提取和检索等问题进行了研究。在运动车辆检测方面,针对目前常用的ViBe算法在检测中出现明显鬼影区域、缓慢目标残影难以消除、检测精确度和鲁棒性不足的问题,本文提出改进,利用灰度信息为像素建立生命长度矩阵,使鬼影或残影快速融入背景样本得以消除。结合最大类间方差法设置自适应阈值,加入良好后处理抑制动态噪音。引入分类算法的统计指标,对车辆检测效果做定性、定量质量评价和分析,实验结果表明,改进算法在较少帧数内去除了鬼影,抑制了运动目标残影,提高了运动车辆检测的整体性能,这为车辆的多目标跟踪和特征提取奠定了良好基础。在车辆多目标跟踪方面,针对目标遮挡、粘连分离,相似物干扰,目标运动不规律影响跟踪稳定性的问题,提出了一种级联分类检测和SVM分类器再识别的区域匹配跟踪算法。在有效提取运动检测得到的目标连通区域的基础上,根据基于HOG特征的级联分类算法有效识别车辆跟踪区域,减少车辆连通域粘连的影响,并且加入基于LBP特征的SVM分类算法二次识别去掉干扰物和相似物,根据区域匹配关联算法保证了跟踪框能够稳定跟踪,通过多组实验验证了本文多目标跟踪算法可以对车辆持续稳定地跟踪,并且具有较高的准确性。在目标特征和检索方面,本文设计了一个基于车辆特征的交通视频检索比对框架,首先对多目标跟踪车辆特征做分析,根据HSV非均匀量化原理提取目标车辆的主区域颜色,利用朴素贝叶斯分类器对车型特征作识别分类。之后将跟踪车辆的特征作结构化存储,同时提出了基于颜色和车型融合的双特征相似车辆检索比对模式,根据倒排索引进行检索比对,快速定位所需要查找的相似车辆。通过实验验证了特征提取和检索比对的有效性和准确性。
[Abstract]:With the development of artificial intelligence, computer vision technology has been integrated into the analysis of intelligent transportation video in intelligent transportation system. The detection of moving targets and the multi-target tracking of vehicles are still facing many challenging problems. At the same time, in more and more vehicle video information, How to do effective feature extraction and convenient and quick retrieval comparison for the vehicles detected and tracked has become an urgent problem in the field of intelligent traffic information. In this paper, moving target detection in traffic video is discussed. The problems of vehicle multi-target tracking and target feature extraction and retrieval are studied in this paper. In the aspect of moving vehicle detection, there are obvious ghost regions in the detection of moving vehicle based on the commonly used ViBe algorithm, so it is difficult to eliminate the residual image of slow target. In this paper, we propose an improved method, in which the gray information is used to build the life length matrix for pixels, so that the ghost or remnant image can be quickly incorporated into the background sample, and the adaptive threshold is set in combination with the maximum inter-class variance method. Adding good post-processing to suppress dynamic noise. Introducing the statistical index of classification algorithm, qualitative, quantitative quality evaluation and analysis of vehicle detection results show that the improved algorithm in less frame number to remove ghost, It can suppress the residual image of moving object, improve the overall performance of moving vehicle detection, which lays a good foundation for vehicle multi-target tracking and feature extraction. In the aspect of vehicle multi-target tracking, the target occlusion, adhesion separation, similar disturbance, etc. In this paper, a region matching tracking algorithm based on cascade classification detection and SVM classifier rerecognition is proposed. The concatenated classification algorithm based on HOG features can effectively identify the vehicle tracking area, reduce the influence of vehicle connectivity, and add the SVM classification algorithm based on LBP feature to remove the interference and similarity. According to the region matching association algorithm, the tracking frame can be tracked stably. The multi-target tracking algorithm in this paper is proved to be able to track vehicles steadily and steadily, and has high accuracy in target feature and retrieval. In this paper, a traffic video retrieval and comparison framework based on vehicle features is designed. Firstly, the multi-target tracking vehicle features are analyzed, and the main region color of the target vehicle is extracted according to the principle of HSV non-uniform quantization. Using naive Bayesian classifier to identify and classify vehicle features, then the tracking vehicle features are stored as structured storage, and a dual-feature similar vehicle retrieval and alignment model based on color and vehicle model fusion is proposed. The efficiency and accuracy of feature extraction and retrieval alignment are verified by experiments.
【学位授予单位】：天津工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】