【作者】 张亮；

【导师】 卢奕南；

【作者基本信息】 吉林大学 ， 计算机应用技术， 2016， 硕士

【摘要】 基于图像或者视频的目标检测和跟踪已经成为计算机视觉领域中的研究热点,目前被广泛应用于视频监控,人机交互,智能检索,车辆导航和蔽障等方面。本文研究的行人检测问题,便是目标检测领域的一个重要分支。本文旨在准确地将目标行人从复杂环境的单一图像或视频图像中检测分离并做进一步的追踪。近年来,行人检测问题由于其广泛的应用和市场价值,受到了越来越多研究者们的关注。但是,由于复杂环境的不确定性以及行人的各异性和多样性,使得行人检测达到精确识别目标仍存在诸多难点,亟待解决。行人检测的传统方法主要分为两类:基于背景建模的方法和基于统计学习的方法。前者由于较低的鲁棒性和较差的抗干扰能力而被研究者所淘汰。因此,现在主流的行人检测算法大多是以统计学习算法为主,该方法首先需要准备训练样本,并从训练样本中提取典型特征以作标注,然后送入分类器中进行学习,最后,根据训练好的分类器进行行人检测。本文对现有的统计学习的方法进行了学习并探究了不同学习算法的优缺点,结合最终建立了一个以改进Hough森林为基本框架的多目标行人检测系统。本文研究的主要目标是实现复杂环境下行人的检测,主要研究内容如下:1)本文研究了基于统计学习的目标检测方法的基本流程,使用统计学习的方法进行行人检测主要需要进行特征提取和机器学习两个工作。2)本文对HOG,LBP,SIFT,Haar-Like四种特征描述子进行了介绍和分析并研究了不同机器学习方法。在Hough算法进行行人检测的研究中,为了更好的描述行人目标,本文采取了一种结合行人Hog特征和Haar-Like特征的复合描述子,该算法能够有效地利用复合特征并结合改进的Hough森林共同形成一套多目标行人检测框架。3)本文选取的Hough森林算法在单一背景下可以实现多目标检测,但是仍然存在精度提升问题,一方面是因为自身的投票机制存在局限性,在复杂背景下会出现较大的识别误差,另一方面是因为投票结果窗口未做处理,导致检测性能下降。因此本文提出了一种基于高斯模板的权重学习方法进行Hough森林优化投票,解决了传统投票机制的局限性问题,并采取窗口融合策略对检测窗口进行进一步处理,最终实现了检测精度的提升。本文主要使用INRIA数据和TUD行人数据库作为行人训练样本参与行人检测系统的训练和学习。实验结果表明本文构建的基于改进Hough森林的多目标行人检测系统在不同的测试数据集下能精确地实现多目标检测,相较于传统的Hog+SVM方法和原始的Hough森林算法具有更优的ROC性能。更多还原

【Abstract】 The research of object detection and tracking in images and videos sequence has become in computer vision field. The use of Object detection and tracking is involved in the applications of the following: video surveillance, human-computer interaction, intelligent retrieval, vehicle navigation and so on. Pedestrian detection of this paper, is an important branch of object detection in the domain of object recognition, which is aimed at detecting, separating and tracking from single image and sequential video frame. In recent years, pedestrian detection has attached more importance of researchers because of its widespread application and deeply important value. But pedestrian detection still has some difficult in precise recognition which is needed to be solved, such as the indeterminacy of environment and the discrepancy of each pedestrian.The traditional method for pedestrian detection is divided into two categories: a method based on the background modeling and a method based on the statistical learning algorithm. The former is not used due to the low robustness and bad antijamming capability. Therefore, most of method for pedestrian detection is based on machine learning. This algorithm first needs to prepare training samples for extracting the features of images. Then the features would be put into the classifier to learn. Finally, we use the learning classifier to detect pedestrians of test images. This paper does some researches on different methods for pedestrian detection and builds a multi-target pedestrian detection system based on modified Hough forest. The main objective of this paper is to solve pedestrians detection and tracking under complicated environment. The main works of this paper are las below:1) Research the method for object detection based on the statistical learning. The method mainly consists of two parts: feature extraction and classifier learning.2) The paper introduces and analyzes four descriptors of feature including HOG, LBP, SIFT and Haar-Like, and researches on different machine learning algorithms as pedestrian classifiers such as SVM algorithm, boosting algorithms, neural network algorithms, etc. We proposes a new pedestrian descriptor combined of Hog feature and Haar-Like feature. Finally, we use Hough forest algorithm as the pedestrian classifier with complex feature because Hough forest can fuse more features to distinguish different categories. The Hough forest can learn and choose the best pedestrian features to form an improved multi-objective pedestrian detection algorithm.3) The original Hough forest algorithm can realize the pedestrian detection in single environment, but it does not apply to complex environment due to limitations of its voting mechanism and raw voting results. Therefore, this paper proposes an optimization algorithm based on Caussian filter area and weight learning to address the limitations of voting module, and adopts a window fusion strategy to process the results, which are together to achieve a lifting detection accuracy.In this paper, we use INRIA dataset and TUD dataset as experimental samples to involve in training and learning module of pedestrian detection system. The results of experiment show that the method proposed in this paper could represent the global information of pedestrians more sufficiently, increase the accuracy of pedestrian detection and satisfy the requirements of real-time detection. And it can get a better ROC performance compared to HOG +SVM algorithm and original Hough forest algorithm.更多还原