基于全景视觉的汽车行驶环境监测系统关键技术研究

发布时间：2018-05-08 15:53

本文选题：鱼眼摄像头 + 全景图像　；参考：《中国农业大学》2017年博士论文

【摘要】：基于全景视觉的汽车行驶环境监测系统可以为驾驶员提供全景成像和目标检测两大驾驶辅助功能,是主动安全领域的重要技术,具有重要的应用价值。现有的目标检测功能主要基于雷达或普通摄像头实现,其检测算法对于大视场、大畸变的鱼眼摄像头并不适用。本文首先对全景成像系统的标定方法进行改进从而快速获得精确的系统参数,然后对基于全景视觉的行人检测算法进行研究。主要研究内容包括:分析了全景成像系统的标定原理。通过对比实验选择折反射模型作为鱼眼摄像头的成像模型,以此为基础进行摄像头的标定和全景成像系统的标定。为鱼眼摄像头设计一种由三个相互垂直的标定板组成的立体标定板,使标定时角点完全覆盖摄像头,从而充分利用鱼眼图像边缘部分以获得更精确的摄像头参数。在全景系统标定板上添加定位块,保证系统标定的稳定性。对鱼眼图像进行前视图投影和直方图均衡化,完成预处理。建立归一化的行人样本库。分别对常用的行人描述特征和机器学习算法进行分析,包括Haar特征结合AdaBoost算法,HOG特征结合SVM算法,以及卷积神经网络进行分类器训练实验,对所得分类器和训练过程进行评价,总结各自的特点。对行人局部HOG特征与整体HOG特征进行对比,提出基于ROI-HOG特征训练SVM分类器;设计卷积神经网络行人检测分类器,优化网络结构参数,并结合无监督CNN提取特征与线性SVM分类器监督学习形成组合分类器,实现快速准确的行人检测。为充分利用鱼眼图像的大视场,研究如何克服图像边缘的大幅度变形。提出了多横摆角前视投影的方法,将一幅鱼眼图像展开为不同横摆角的前视投影图,然后进行行人检测。图像任意位置的行人在某一范围内的虚拟横摆角下的前视图中,都可去除仿射变形恢复正常人体比例,方便后续进行行人检测。通过实验总结前视图横摆角设置规则,尽量减少一幅鱼眼图像的展开数,减少检测耗时。进行实车全景系统标定实验,结果表明系统具有较高的标定精度和标定稳定性。采集行车视频数据,逐帧标记行人形成测试集,在PC上设计评估软件,运行行人检测程序,评估行人检测算法性能,结果表明在一定距离范围内本文算法可以实现较高的行人检测率。
[Abstract]:The vehicle driving environment monitoring system based on panoramic vision can provide driving assistant functions of panoramic imaging and target detection. It is an important technology in active safety field and has important application value. The existing target detection function is mainly based on radar or ordinary camera, and its detection algorithm is not suitable for large field of view and large distortion fish-eye camera. In this paper, the calibration method of panoramic imaging system is improved to obtain the accurate system parameters quickly, and then the pedestrian detection algorithm based on panoramic vision is studied. The main contents are as follows: the calibration principle of panoramic imaging system is analyzed. The refraction model is chosen as the imaging model of fish-eye camera by contrast experiment, and the camera and panoramic imaging system are calibrated based on the model. A stereo calibration board composed of three vertical calibration boards is designed for the fish-eye camera, which can cover the camera completely when the corners are calibrated, thus making full use of the edge part of the fish-eye image to obtain more accurate camera parameters. A positioning block is added to the panoramic system calibration board to ensure the stability of the system calibration. The front view projection and histogram equalization are used to preprocess the fish eye image. A normalized pedestrian sample bank is established. The common pedestrian description features and machine learning algorithms are analyzed respectively, including Haar feature and AdaBoost algorithm combined with SVM algorithm, and convolution neural network for classifier training experiment. The classifier and training process are evaluated. Summarize their own characteristics. By comparing the local HOG features of pedestrians with the overall HOG features, a SVM classifier based on ROI-HOG feature training is proposed, and a pedestrian detection classifier based on convolution neural network is designed to optimize the network structure parameters. Combined with unsupervised CNN feature extraction and linear SVM classifier supervised learning, a combined classifier is formed to realize fast and accurate pedestrian detection. In order to make full use of the large field of view of the fish-eye image, this paper studies how to overcome the large deformation of the image edge. A method of forward projection with multiple yaw angles is proposed. A fish-eye image is expanded into a forward projection image with different yaw angles, and then pedestrian detection is carried out. In the front view of a virtual yaw angle in any position of the image, the affine deformation can be removed to restore the normal proportion of the human body, and it is convenient to carry out pedestrian detection. The rules of yaw angle setting in front view are summarized by experiments to reduce the expansion number of a fish-eye image and the detection time. The calibration experiment of real vehicle panoramic system shows that the system has high calibration accuracy and stability. Collect the video data of the vehicle, mark the pedestrian to form the test set frame by frame, design the evaluation software on the PC, run the pedestrian detection program, evaluate the performance of the pedestrian detection algorithm, The results show that the proposed algorithm can achieve high pedestrian detection rate within a certain range of distances.
【学位授予单位】：中国农业大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：U463.6

【参考文献】