基于主动学习的车载单目视觉车辆检测与跟踪研究

发布时间：2018-06-17 19:37

本文选题：视觉车辆检测 + 视觉车辆跟踪　；参考：《中国科学技术大学》2017年硕士论文

【摘要】：随着汽车保有量的迅猛增长,公路交通事故已经成为全球范围内日趋严重的公共安全问题,亟待解决。前碰撞预警系统是智能辅助驾驶系统的重要组成部分,能有效降低公路交通事故发生的概率。车辆检测和跟踪的准确性、连续性和实时性是影响该系统功能发挥的决定性因素。其中,车辆定位的准确性和连续性是预警功能的前提,而实时性是预警功能有效发挥的关键,能使驾驶者及早发现险情。因此本文致力于车载单目视觉的车辆检测与跟踪算法研究,具体研究内容如下:基于主动学习的分类器模型训练。基于机器学习的视觉车辆检测需要大量带有标签的样本数据,用以训练出能够准确分类图像中车辆与背景的分类器模型。本文提出一种基于错误分类样本抽样策略的主动学习算法,以较小的人工标注成本获得最具信息量的样本数据,迭代训练优化分类器的性能。Adaboost(Adaptive Boosting)级联多目标车辆检测。为了提高车辆检测的准确性,本文提出一种分区域多分类器车辆检测方法。根据车辆特征在检测视野中的差异,把待检测车辆分类为前向车辆、左斜侧向车辆和右斜侧向车辆,分别训练级联分类器进行检测。同时,为了提高车辆检测速度,提出一种结合相机标定的多分辨率加速车辆检测算法,对检测视野中远近不同的车辆采用不同程度的图像降采样分别检测。HOG(Histogram of Oriented Gradients)特征跟踪与 Adaboost 检测融合。针对Adaboost级联车辆检测结果不够连续的问题,提出一种Adaboost级联检测与HOG特征跟踪相互融合的车辆检测跟踪算法。通过HOG特征跟踪的融入,提高了约10%的车辆检测率,使检测结果更加连续。前碰撞预警系统设计实现。文章最后应用本文研究的车辆检测跟踪算法设计出一套前碰撞预警系统,通过真实交通场景测试,该系统可以实时、准确和连续的检测跟踪前方车辆并计算与其距离,实时监控前方潜在的碰撞危险,及时发出预警信号,从而避免交通事故的发生。
[Abstract]:With the rapid growth of vehicle ownership, road traffic accidents have become more and more serious public safety problems all over the world. Pre-collision warning system is an important part of intelligent auxiliary driving system, which can effectively reduce the probability of road traffic accidents. The accuracy, continuity and real-time of vehicle detection and tracking are the decisive factors affecting the function of the system. Among them, the accuracy and continuity of vehicle positioning is the premise of early warning function, and real-time is the key to the effective use of early warning function, which can enable the driver to detect the danger as early as possible. Therefore, this paper focuses on vehicle detection and tracking algorithm based on vehicle monocular vision. The research contents are as follows: classifier model training based on active learning. Visual vehicle detection based on machine learning requires a large number of labeled sample data to train a classifier model that can accurately classify vehicles and backgrounds in images. In this paper, an active learning algorithm based on sample sampling strategy for error classification is proposed to obtain the most informative sample data at a lower cost of manual annotation, and iterative training to optimize the performance of the classifier. Adaboosting Adaptive boost) cascade multi-objective vehicle detection. In order to improve the accuracy of vehicle detection, this paper presents a multi-classifier vehicle detection method. According to the difference of vehicle characteristics in the field of vision, the vehicles to be tested are classified as forward vehicles, left oblique vehicles and right oblique vehicles, and the cascaded classifiers are trained for detection. At the same time, in order to improve the speed of vehicle detection, a multi-resolution accelerated vehicle detection algorithm combined with camera calibration is proposed. Different degrees of image demotion were used to detect different vehicles in the visual field. The feature tracking of the histogram of oriented radientsand the fusion of Adaboost detection were used respectively. In order to solve the problem that the detection results of Adaboost cascaded vehicles are not continuous, a vehicle detection and tracking algorithm based on Adaboost cascade detection and hog feature tracking is proposed. By means of HOG feature tracking, the vehicle detection rate is increased by about 10%, and the detection results are more continuous. Design and implementation of pre-collision warning system. Finally, using the vehicle detection and tracking algorithm studied in this paper, a pre-collision warning system is designed. Through the real traffic scene test, the system can detect and track the vehicle in front of the vehicle in real time, accurately and continuously, and calculate the distance between the vehicle and the vehicle. Real-time monitoring of potential collision hazards ahead, timely warning signals to avoid traffic accidents.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U463.6;TP391.41

【参考文献】