基于单目视觉的前方车辆检测与跟踪方法研究

发布时间：2018-04-23 06:13

本文选题：安全辅助驾驶系统 + 车辆检测　；参考：《合肥工业大学》2015年硕士论文

【摘要】：随着公路交通事业的迅速发展,给人们的生活带来便利的同时,也造成了道路交通安全问题日益突出。安全辅助驾驶系统作为智能交通系统的重要组成部分,能够有效地预防交通事故的发生,提高行车的安全性。而前方车辆检测与跟踪是安全辅助驾驶系统的核心环节,为车辆的信息提取以及行为分析提供了重要的保证。本文在分析和比较国内外各种算法的基础上,研究并提出了相应的基于单目视觉的车辆检测与跟踪算法。全文主要内容如下：1)对前方车辆检测与跟踪的研究背景和意义进行了探讨,阐述了常见的基于单目视觉的前方车辆检测与跟踪方法,并分析了这些方法的优缺点,为后续研究作准备。2)研究了基于AdaBoost的前方车辆检测算法。选取Haar-like特征作为图像特征,利用Gentle AdaBoost算法和CasCade算法对训练样本进行离线学习,得到级联结构的车辆分类器；检测过程中,采用等比放大检测窗口的方式扫描待检图像,利用车辆分类器对检测窗口进行分类,最后综合各个检测窗口的结果,得出车辆的最终位置。实验结果表明,该方法能够有效地检测出前方车辆,具有一定的鲁棒性,基本上满足实时性要求。3)提出一种基于改进TLD的前方车辆跟踪算法。TLD算法是一种新颖的目标跟踪算法,在给定极少的先验知识的情况下,能够迅速地学习目标特征并进行有效的跟踪。而车辆的先验知识可由前方车辆检测算法提供,因此,TLD跟踪算法完全能够适用前方车辆跟踪问题上。然而,TLD跟踪模块均匀地选取特征点进行跟踪,无法保证所选特征点被可靠地跟踪。针对这个问题,提出一种基于关键特征点的选取方式,保证所选特征点能够被正确可靠地跟踪,防止跟踪发生漂移,提高跟踪精度。另一方面,在TLD检测模块中引入了基于轨迹连续性的在线位置预测,在保证正确跟踪的前提下,缩小了检测范围,提高了运算速度。最后,利用改进的TLD算法对前方车辆进行跟踪。实验结果表明,该算法能够有效的对前方车辆跟踪,且在各种较难处理的情况下具有较好的跟踪效果。
[Abstract]:With the rapid development of highway traffic, it brings convenience to people's life, but also causes the problem of road traffic safety more and more prominent. As an important part of intelligent transportation system, safety assistant driving system can effectively prevent traffic accidents and improve the safety of traffic. The detection and tracking of vehicle in front is the core of the safety assistant driving system, which provides an important guarantee for the information extraction and behavior analysis of the vehicle. Based on the analysis and comparison of various algorithms at home and abroad, this paper studies and proposes the corresponding vehicle detection and tracking algorithm based on monocular vision. The main contents of this paper are as follows: (1) the research background and significance of forward vehicle detection and tracking are discussed, and the common methods of forward vehicle detection and tracking based on monocular vision are expounded, and the advantages and disadvantages of these methods are analyzed. The forward vehicle detection algorithm based on AdaBoost is studied. Haar-like feature is selected as image feature, Gentle AdaBoost algorithm and CasCade algorithm are used to study the training sample offline, and the cascade structure vehicle classifier is obtained. In the process of detection, the image is scanned by equal ratio amplification detection window. The detection window is classified by vehicle classifier and the final position of the vehicle is obtained by synthesizing the results of each detection window. The experimental results show that the proposed method can detect the vehicle in front effectively, which is robust, and basically meets the real-time requirement. 3) A novel target tracking algorithm based on improved TLD is proposed. Given minimal prior knowledge, we can quickly learn target features and track them effectively. The prior knowledge of the vehicle can be provided by the forward vehicle detection algorithm, so the TLD tracking algorithm is fully applicable to the forward vehicle tracking problem. However, the TLD tracking module selects the feature points uniformly and can not guarantee that the selected feature points can be tracked reliably. To solve this problem, a method based on key feature points is proposed to ensure that the selected feature points can be tracked correctly and reliably, to prevent the drift of tracking, and to improve the tracking accuracy. On the other hand, the on-line position prediction based on trajectory continuity is introduced into the TLD detection module, which reduces the detection range and improves the operation speed on the premise of correct tracking. Finally, the improved TLD algorithm is used to track forward vehicles. The experimental results show that the algorithm can track the vehicle in front effectively and has a good tracking effect under various difficult cases.
【学位授予单位】：合肥工业大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：U495;TP391.41

【相似文献】