基于雷达与图像信息融合的路面目标识别与应用

发布时间：2018-06-26 03:21

本文选题：多传感器 + 栅格地图　；参考：《南京理工大学》2017年硕士论文

【摘要】：无人驾驶车辆在军用和民用领域都具有重要的研究价值和广泛的应用前景。道路环境感知技术是无人自主车的关键技术之一,多传感器由于其性能互补,能获得更准确、更高效的感知效果。本文旨在对多线激光雷达数据与图像数据进行融合,发掘这两种传感器的有效信息,从而对无人驾驶车辆的路面目标进行识别。完成的工作主要如下:首先基于32线激光雷达进行路面障碍物的初步检测,对雷达数据构建了基于相对高度差信息的栅格地图,然后提出了一种基于雷达数据点距离信息预判断的密度聚类算法,标记出雷达数据点密集的区域,也就是障碍物区域。通过先验几何知识过滤掉非路面障碍。实验结果表明,该方法能够成功检测路面障碍物,并获知障碍物的距离信息。研究基于图像信息的方法对路面行人进行检测。首先用混合高斯模型对背景进行建模,过滤大部分背景,再采用形态学滤波算法的腐蚀和膨胀,剔除孤立的小点,融合邻近的数据点,最后采用矩形包围轮廓,找出障碍物的位置。比较了基于图像信息的行人检测算法与基于激光雷达的障碍检测算法,分析这两种传感器的优缺点。多传感器的信息融合建立在各数据坐标统一的基础上。首先对雷达进行参数标定,根据标定物的角点对应,求出雷达数据点由原始的极坐标形式转换到世界坐标的转换矩阵。然后采用Tsai方法对摄像头进行标定,最后得到雷达坐标系到图像像素坐标系的投影矩阵。提出基于障碍物的数据融合方法,将雷达中检测到的障碍物的深度信息与图像数据融合,使图像的像素点不仅包含RGB信息,也包含距离信息,为下一步识别目标以及确定目标的距离是否安全作准备。确认路面障碍物,特别是行人的位置对无人驾驶车辆的安全驾驶具有重要的意义。将经雷达初步检测后得到的障碍物感兴趣区域投影到图像上,对图像上的感兴趣区域提取HOG特征算子。将行人库数据作为正样本,不包含行人的无人车真实行驶场景作为负样本,训练线性SVM分类器,识别障碍物是否为行人,并根据融合后的像素点的深度信息判断障碍物的距离。实验证明,该方法能实现实时路面目标识别,能应用到无人车的自主导航系统中,具有实际应用意义。
[Abstract]:Driverless vehicles have important research value and wide application prospect in both military and civil fields. Road environment sensing technology is one of the key technologies of unmanned vehicles. Because of its complementary performance, multi-sensors can obtain more accurate and efficient sensing effect. The purpose of this paper is to fuse the multi-line lidar data with the image data, to discover the effective information of the two sensors, and to identify the road targets of driverless vehicles. The main work is as follows: firstly, based on 32-line lidar, the initial detection of road obstacles is carried out, and the raster map based on the relative height difference information is constructed for the radar data. Then, a density clustering algorithm based on the pre-judgment of radar data point distance information is proposed, which marks the area where radar data points are dense, that is, obstacle area. The non-road barrier is filtered out by a prior knowledge of geometry. The experimental results show that the method can successfully detect obstacles and obtain the distance information of obstacles. A method based on image information for pavement pedestrian detection is studied. First, the background is modeled by mixed Gao Si model, most of the background is filtered, then the corrosion and expansion of morphological filtering algorithm is adopted, the isolated small points are eliminated, the adjacent data points are fused, and the outline is surrounded by a rectangle. Find out the position of the obstacle. The pedestrian detection algorithm based on image information and the obstacle detection algorithm based on lidar are compared, and the advantages and disadvantages of the two sensors are analyzed. The multi-sensor information fusion is based on the unification of the data coordinates. First, the radar parameters are calibrated, and the transformation matrix of radar data points from the original polar coordinate to the world coordinate is obtained according to the corner point correspondence of the calibrated object. Then the camera is calibrated by Tsai method, and the projection matrix from radar coordinate system to image pixel coordinate system is obtained. A method of data fusion based on obstacle is proposed. The depth information of obstacle detected in radar is fused with image data, so that the pixel of image contains not only RGB information, but also distance information. Prepare for the next step to identify the target and determine whether the target's distance is safe. It is important to confirm the position of obstacles, especially pedestrians, for the safe driving of driverless vehicles. The region of interest obtained by radar preliminary detection is projected onto the image, and the Hog feature operator is extracted for the region of interest on the image. Taking the pedestrian bank data as the positive sample and the actual driving scene without pedestrians as the negative sample, the linear SVM classifier is trained to identify whether the obstacle is a pedestrian, and the distance of the obstacle is judged by the depth information of the fused pixel. Experiments show that this method can realize real-time road surface target recognition and can be applied to autonomous navigation system of unmanned vehicle.
【学位授予单位】：南京理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U463.6;TN958.98

【相似文献】