多分类器级联的街道场景遮挡行人检测

发布时间：2018-02-27 00:01

本文关键词： 行人检测遮挡 HOG 多分类器 BING　出处：《南昌航空大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着人工智能和深度学习技术的不断发展,目标检测、识别、跟踪等计算机视觉应用已经越来越多地出现在我们的生活之中。行人检测由于其应用领域的广泛,近年来得到了学者们的广泛关注。由于行人自身的非刚性以及所处环境的复杂性,各类场景下的行人检测依旧是目标检测领域的难点与重点。本文首先介绍了行人检测领域的研究背景与意义,结合具体事例阐述了行人检测在实际生活中的各项应用。然后介绍了目前行人检测领域的技术难点与国内外研究现状,并对常用的特征算子与分类方法做了详细的介绍。本文的核心工作是针对日常街道场景中各类不同遮挡状况的行人,提出一种多分类器级联检测的方式,并通过训练通用行人目标检测BING(binarized normed gradients)模板代替滑动窗口扫描来加速检测过程。本文的主要工作可以分为以下五个方面:1、基于HOG(histogram of oriented gradient)特征和线性SVM(Support Vector Machines)分类器,从INRIA数据集挑选合适的正负样本训练了一个无遮挡行人分类器。该分类器用于研究行人不同部位被遮挡对检测效果的影响是否不同。验证实验结果的数据一部分来自CVC-05 part occlusion数据集,一部分来自自己收集制作的遮挡行人数据集。2、基于INRIA、CVC05-PartOcclusion数据集和自己收集的行人图片制作了一系列遮挡行人专用正样本,并利用这些数据集训练了九个检测不同部位被遮挡的行人分类器。其中无遮挡行人分类器一个,腿部遮挡行人分类器两个,左右半身遮挡行人分类器各三个。3、通过级联的方式加速多分类器检测过程。分类器设置为两级,其中第一级一个分类器,第二级八个分类器,两级分类器之间串联,第二级分类器之间并联。只有检测得分大于第一级分类器设置的阈值的检测窗口才会输入到第二级分类器做进一步的检测。通过级联分类器的方式使得检测时间与单分类器基本持平。4、在多分类器检测结果融合阶段提出一种NMS+Merging的方式。首先对于同一分类器的所有检测窗口采取NMS的方式去除得分低的检测窗口,然后对于不同分类器保留下来的检测窗口采取一种两两融合的方式进而得到最终的检测结果。5、针对目标检测领域滑动窗口检测法提取候选窗口数量冗余的问题提出一种基于BING算法改进的通用行人目标快速提取算法。采用Caltech数据集训练,并根据行人特殊宽高比设置行人检测专用BING模板。该方法可以在保持原有检测精度的同时进一步缩短检测阶段所消耗的时间。
[Abstract]:With the development of artificial intelligence and depth learning technology, the application of computer vision, such as target detection, recognition, tracking and so on, has been more and more popular in our life. In recent years, scholars have paid more and more attention to it. Due to the nonrigid nature of pedestrians and the complexity of their environment, Pedestrian detection in all kinds of scenarios is still a difficult and important point in the field of target detection. Firstly, this paper introduces the background and significance of pedestrian detection. In this paper, the application of pedestrian detection in real life is expounded with concrete examples. Then, the technical difficulties in the field of pedestrian detection and the current research situation at home and abroad are introduced. The main work of this paper is to propose a multi-classifier cascade detection method for pedestrians with different occlusion conditions in street scenes. The detection process is accelerated by training the general pedestrian target detection BING(binarized normed gradientstemplate instead of sliding window scanning. The main work of this paper can be divided into the following five aspects: 1, based on HOG(histogram of oriented gradient-based feature and linear SVM(Support Vector machines classifier. An unoccluded pedestrian classifier is trained by selecting the appropriate positive and negative samples from the INRIA data set. The classifier is used to study whether the influence of different parts of the pedestrian on the detection effect is different. From the CVC-05 part occlusion dataset, Some of them are collected and made by themselves. 2. Based on the INRIAA CVC05-PartOcclusion dataset and the pedestrian images collected by them, a series of special positive samples of occluded pedestrians are made. Using these data sets, nine pedestrian classifiers are trained to detect the occlusion of different parts, including one unoccluded pedestrian classifier and two leg occluded pedestrian classifiers. The pedestrian classifiers, three from the left and the other half, speed up the process of multi-classifier detection by cascading. The classifier is set to two levels, one classifier in the first stage, eight classifiers in the second stage, and the other in series between the two classifiers. Only a detection window with a detection score greater than the threshold set by the first level classifier will be input to the second level classifier for further detection. The detection time is made by cascading classifiers. At the fusion stage of multi-classifier detection results, a NMS Merging method is proposed. Firstly, NMS is used to remove the low-score detection window for all detection windows of the same classifier. Then, for the detection window retained by different classifiers, a pairwise fusion method is adopted to obtain the final detection result .5. the problem of extracting redundant number of candidate windows by sliding window detection method in target detection field is raised. This paper presents an improved fast pedestrian target extraction algorithm based on BING algorithm. Caltech data set is used to train the pedestrian target. According to the special aspect ratio of pedestrians, a special pedestrian detection BING template is set up. This method can further shorten the time consumed in the detection stage while maintaining the original detection accuracy.
【学位授予单位】：南昌航空大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】