逐步求精的分层汽车检测框架研究与实现

发布时间：2018-07-07 16:57

本文选题：背景差分 + Adaboost　；参考：《中国地质大学(北京)》2017年硕士论文

【摘要】：当前,为了对日益恶化的交通情况进行更好的管理,在道路上安装了越来越多的摄像机,形成了海量的视频数据。因此,研究如何快速准确的从视频中提取车辆目标意义重大。本文首先探讨了目标检测的基本理论和方法,以及国内外在该领域的研究现状,然后探讨了最新的研究成果。其中基于深度学习的目标检测方法,达到了很好的检测效果,逐渐成为该领域的研究热点。本文的主要工作围绕将各种车辆目标检测方法进行结合的研究,内容如下:(1)基于背景差分的目标检测算法:基于背景差分检测算法是最早的目标检测算法,它具有速度快,漏检率低的特点,算法设计简单高效,应用广泛。早期的目标检测算法几乎全部是基于背景差分算法建立的。但是背景差分算法有一些缺陷,比如无法有效的克服阴影的误检问题。(2)基于Adaboost的目标检测算法:基于Adaboost的目标检测算法是后来发展起来的目标检测算法,该算法首先在人脸检测上成功应用,是第一个将人脸检测做到实时的算法,并且有效的克服了光照和姿势变化的影响,但是该算法的计算复杂度相对较高,误检较多,导致不能在很多嵌入式设备上应用。(3)基于深度神经网络的图像识别方法:近年来,深度神经网络掀起了人工智能的革命,将图像识别的准确率提升到新的高度,比如在人脸识别准确率方面已经超过人类。但是计算复杂度非常高,一般需要GPU才可以做到实时性。(4)将背景差分,Adaboost和深度神经网络进行级联的目标检测算法:很多实验表明,背景差分算法计算复杂度低,并且能够高效的检测出前景区域,因此提出先对图像进行背景差分,排除背景区域,只在前景区域上运行Adaboost目标检测算法,最后再针对Adaboost输出的目标结果使用深度神经网络进行二次识别,过滤掉误检的车辆目标。这样可以很好的利用上面各种算法的优点,又避免了各自的缺陷。同时,本文对Adaboost目标检测算法又进行了改进,即通过标定出图像不同位置的目标尺寸以减少Adaboost扫描窗口的数量。
[Abstract]:At present, in order to better manage the deteriorating traffic situation, more and more cameras are installed on the road, forming a huge amount of video data. Therefore, it is significant to study how to extract vehicle targets from video quickly and accurately. In this paper, the basic theories and methods of target detection are discussed, and the research status in this field at home and abroad is also discussed, and then the latest research results are discussed. Among them, the target detection method based on depth learning has achieved a good detection effect, and has gradually become a research hotspot in this field. The main work of this paper is focused on the research of combining various vehicle target detection methods. The contents are as follows: (1) background differential detection algorithm is the earliest object detection algorithm, and it has fast speed. Low detection rate, simple and efficient algorithm design and wide application. Almost all of the early target detection algorithms were based on background differential algorithm. But the background difference algorithm has some defects, such as can not effectively overcome the shadow false detection problem. (2) Target detection algorithm based on Adaboost: the target detection algorithm based on Adaboost is a target detection algorithm developed later. First, the algorithm is successfully applied to face detection, and it is the first one to achieve real-time face detection, and it effectively overcomes the influence of illumination and pose changes, but the computational complexity of the algorithm is relatively high, and the error detection is more. As a result, it can not be applied to many embedded devices. (3) Image recognition method based on depth neural network: in recent years, depth neural network has set off a revolution of artificial intelligence, which raises the accuracy of image recognition to a new height. For example, the accuracy of face recognition has surpassed that of humans. However, the computational complexity is very high, GPU is generally required to achieve real-time. (4) the background differential Adaboost and the depth neural network are cascaded to detect the target. Many experiments show that the background differential algorithm has a low computational complexity. And the foreground region can be detected efficiently, so the background difference is proposed, the background area is excluded, and only the Adaboost target detection algorithm is run on the foreground region. In the end, a depth neural network is used for the secondary recognition of the target output from Adaboost to filter out the vehicle target that has been misdetected. In this way, we can make good use of the advantages of the above algorithms, and avoid their own defects. At the same time, the algorithm of Adaboost target detection is improved, that is, to reduce the number of Adaboost scanning windows by calibrating the target size at different positions of the image.
【学位授予单位】：中国地质大学(北京)
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【相似文献】