基于特征提取的高速公路隧道环境下行人检测研究

发布时间：2018-03-09 21:03

本文选题：隧道　切入点：行人检测　出处：《昆明理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：隧道区域是高速公路管理的重点区域,行人和非机动车辆违规进入高速公路隧道内会严重影响高速公路的正常运行,造成巨大的安全隐患。因此,针对隧道环境下视频监控中的行人检测技术是高速公路正常运营的重要保障。隧道环境内,环境光照条件差,在图像中产生大量噪声,行人在隧道内目标小,像素低,给隧道环境下行人检测带来很大挑战。本文主要研究了视频检测中的前景目标与背景目标的分割方法,使用了基于数学特征提取方法与卷积神经网络的行人目标检测方法。并且针对提取特征训练的分类器遍历搜索慢,在隧道场景下采用运动信息缩小搜索范围,节省了搜索时间。另外针对隧道环境下噪声大行人特征提取困难的问题,利用卷积神经网络对特征提取的优势特点,训练了端到端的隧道场景下行人检测网络。本文的主要研究内容如下:(1)一般的行人检测分类器的训练中通常采用单一的HOG特征,在隧道环境下检测准确率偏低。本文通过引入一种局部二值模式(LBP)特征与梯度方向直方图特征(HOG)串联输入到支持向量机的分类模型中,训练得到的基于联合特征行人检测器大幅提升了隧道环境下行人检测的准确率。(2)基于HOG特征与LBP特征串联训练的分类器一般采用滑动窗口遍历搜索整个图像的策略,这样造成了巨大的时效性损失。在高速公路隧道中,监控画面出现在固定场景下。根据视频监控的这一特点,通过提取行人移动信息,将分类器检测与一种改进的高斯混合背景差分方法相结合,提取图像中运动区域,减少分类器对图像的搜索次数,大幅提升了算法系统对行人的识别效率。(3)针对高速公路隧道环境噪声造成行人与环境轮廓边界弱,传统机器学习方法难以提取有效特征的问题,本文利用卷积神经网络高效的特征提取能力,通过改进候选框提取方法,使用RPN候选框提取网络,在选用单幅图片候选框少的情况下训练出行人检测的单一目标识别网络。对候选框提取网络与行人检测网络进行了训练,得到端到端的行人检测网络。相对于特征设计的行人检测模型,大幅度的提升隧道环境行人检测的准确率,且在一定程度上提升基于RCNN算法框架下的行人检测速度。针对高速公路隧道环境下行人检测的要求,研究了基于特征提取的分类器模型对隧道应用场景的适应性,并提出相应的方法改进检测模型。将区域卷积神经网络应用在氋速公路隧道场景下的行人检测,同时训练了端到端的深度行人检测模型。提升了隧道监控场景下行人检测的准确率,对基于卷积神经网络的其他目标物识别工作具有一定的借鉴意义。
[Abstract]:Tunnel area is the key area of highway management. Illegal entry of pedestrians and non-motorized vehicles into expressway tunnel will seriously affect the normal operation of expressway and cause huge safety hazard. The pedestrian detection technology in video surveillance in tunnel environment is an important guarantee for the normal operation of highway. In the tunnel environment, the environment lighting conditions are poor, a lot of noise is produced in the image, the pedestrian in the tunnel has small targets and low pixels. It brings great challenge to pedestrian detection in tunnel environment. This paper mainly studies the segmentation method of foreground target and background object in video detection. The pedestrian target detection method based on mathematical feature extraction and convolution neural network is used. In addition, aiming at the difficult problem of extracting noisy pedestrian features in tunnel environment, we use convolutional neural network to extract features. The downlink detection network of end-to-end tunnel scene is trained. The main contents of this paper are as follows: 1) in the training of pedestrian detection classifier, a single HOG feature is usually used. In this paper, a local binary pattern (LBP) feature and gradient direction histogram feature (hog) are introduced into the classification model of support vector machine in series. The trained pedestrian detector based on joint feature greatly improves the accuracy of pedestrian detection in tunnel environment. (2) the classifier based on HOG feature and LBP feature series training generally uses sliding window traversal strategy to search the whole image. This results in a huge loss of timeliness. In highway tunnels, surveillance images appear in fixed scenes. According to this characteristic of video surveillance, by extracting pedestrian movement information, Combining the classifier detection with an improved Gao Si mixed background differential method, the moving region of the image is extracted, and the search times of the image are reduced. The efficiency of pedestrian recognition in the algorithm system is greatly improved. (3) aiming at the problem that the boundary between pedestrian and environment is weak due to the noise in highway tunnel environment, the traditional machine learning method is difficult to extract effective features. In this paper, we use convolutional neural network to extract features, improve the method of candidate extraction, and use RPN candidate to extract the network. The single target recognition network is trained under the condition of few single image candidate frames, and the candidate extraction network and pedestrian detection network are trained. Get the end to end pedestrian detection network. Compared with the feature designed pedestrian detection model, greatly improve the accuracy of pedestrian detection in tunnel environment. To some extent, the speed of pedestrian detection based on RCNN algorithm is improved. According to the requirement of pedestrian detection in highway tunnel environment, the adaptability of classifier model based on feature extraction to tunnel application scene is studied. The corresponding method is put forward to improve the detection model. The regional convolution neural network is applied to pedestrian detection in the scene of highway tunnel. At the same time, it trains the end-to-end depth pedestrian detection model, improves the accuracy of downlink detection of tunnel monitoring scene, and has some reference significance for other target recognition work based on convolution neural network.
【学位授予单位】：昆明理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U458;TP391.41

【参考文献】