夜间车辆跟踪与自动评价技术

发布时间：2018-03-02 06:13

本文关键词： 夜间车辆检测阈值分割 Haar特征车辆跟踪自动评价　出处：《北京交通大学》2014年硕士论文　论文类型：学位论文

【摘要】：车辆检测和跟踪是智能交通的一个重要内容,也为“十二五规划”智慧城市建设的首要目标——平安城市提供关键技术。该领域涉及的运动目标检测和多目标跟踪技术一直是计算机视觉和机器学习的研究热点。本文在前期调研工作的基础上,基于车灯特征对夜间车辆进行检测,并训练分类器来降低虚警率,在此基础上实现鲁棒的夜间车辆跟踪。利用自动评价技术对夜间车辆跟踪进行自动评价,并根据结果提出了改进的方案。本文的主要工作有： 1、基于车灯特征的车辆检测技术的研究。本文首先对比了多种车辆检测和跟踪的技术,例如帧间差分法、背景差分法和自适应阈值分割法。针对这些传统方法只适用于白天、光照充足条件下的车辆检测和跟踪而不适用于夜间的问题,提出了基于车灯的车辆检测。在训练阶段,采用针对车灯的Haar特征并融合多尺度的几何、形状特征进行Adaboost分类器训练。在测试阶段,根据统计直方图设定阈值,分割夜间图像中的车灯。所有检测到的车灯通过Adaboost分类器分类出正样本,在此基础上根据速度、几何、形状等特征的相似性建立数据关联进行多目标跟踪。 2、针对车辆跟踪的自动评价技术。对车辆跟踪效果的评价是一个工作量极大的部分,为此,本文提出了一种自动评价的方法来对车辆跟踪的各项性能进行自动的评价。首先是介绍了一款半自动标注工具,来获取真实的车辆跟踪结果。利用真实的车辆跟踪结果和车辆跟踪算法生成的跟踪结果通过自动评价技术自动生成各项评价指标。这些评价指标包括正确率、缺失率、误判率和转变率。各项评价指标的数据显示,本文提出的方法能够对车辆进行实时鲁棒的跟踪。
[Abstract]:Vehicle detection and tracking is an important part of the intelligent transportation system, but also for the primary objective of "12th Five-Year plan" wisdom City Construction -- safe city to provide key technology. Tracking technology of moving target detection in the fields and multi targets has been a research focus of computer vision and machine learning. This paper based on the previous research work. On the night, vehicle detection and classifier training based on the characteristics of light, to reduce the false alarm rate, to achieve robust on the basis of the night. On the night of the vehicle tracking vehicle tracking automatic evaluation using automatic evaluation technology, and put forward the improvement scheme according to the results.
The main work of this article is as follows:
1, study the characteristics of vehicle detection technology based on. This paper compares a variety of vehicle detection and tracking technology, such as the inter frame difference method, background difference method and adaptive threshold segmentation method. The traditional method is only suitable for daytime, adequate light under the condition of vehicle detection and tracking is not suitable for the night, the lights of the vehicle detection based on. During the training phase, the geometry for Haar lights and feature fusion of multi-scale, shape feature Adaboost classifier training. During the testing phase, according to the statistics histogram set the threshold segmentation in the image, the night lights. All the detected light by Adaboost classifier classification the positive samples, based on the speed, geometry, shape similarity and other characteristics of the establishment of Data Association for multiple target tracking.
2, according to the automatic evaluation technology of vehicle tracking. To evaluate the effect of vehicle tracking is a great part of a workload for this purpose, this paper presents a method for automatic evaluation to the evaluation of the performance of automatic vehicle tracking. At first it introduces a semi-automatic annotation tool to obtain real tracking results the use of vehicles. Vehicle tracking and generate results of real vehicle tracking algorithm based on tracking results through the automatic evaluation technology of automatic generation of evaluation indexes. These indexes include the correct rate, loss rate, error rate and transformation rate. The evaluation index data show that the proposed method can real-time robust vehicle tracking.

【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：U495;TP391.41

【参考文献】