基于激光与视觉信息融合的运动目标检测与场景重建方法研究

发布时间：2018-04-20 22:13

本文选题：激光与视觉信息融合 + 运动目标检测　；参考：《中国科学技术大学》2017年硕士论文

【摘要】：运动目标检测与场景重建作为视频监控和三维地图构建的关键技术,是进行实时导航避障与路径规划的基础。它们作为环境感知任务的子问题,与机器人、无人机、无人车、体感游戏等领域的发展密不可分,是人类智能生活的一部分。当前对该问题的研究多基于单一的视觉或激光传感器,受限于其视野范围、数据量、数据丰富度以及实时性和抗干扰性等局限,往往难以满足实时的多任务场景需求。本文以激光信息和视觉信息的相互补充与约束,设计了一个可以同时进行实时运动目标检测与场景重建的感知系统,并对上述两个子问题进行了深入研究。主要内容如下:1、对比分析了多传感器硬件平台Multisense-SL中激光与视觉传感器在各项性能上的差异,论证了两者的融合在动态场景感知问题上的合理性;并基于ROS软件平台的消息发布与订阅模式,设计了可以同时进行多项实时感知任务的融合框架。2、针对基于视觉的背景减除法进行的运动目标检测中的"背景显露区域"误判问题和三维前景提取的不完整性问题,提出了一种特征层下的融合运动目标检测算法。以基于视觉分拣出的激光前景点为启发信息,进行二维图像邻域搜索下的三维距离约束的前景聚类,同时得到了二维图像前景图和三维空间前景点云。在完整提取前景对象的同时有效的抑制了 "背景显露区域"的误判。3、针对激光与视觉数据底层融合的噪音问题和直接通过传感器间外标定关系配准的点云"失配现象",在进行场景的融合重建中进行了优化。在预处理阶段提出了视场约束、时效约束和遮挡约束,有效地滤除了绝大部分杂点;并在融合优化阶段提出了基于聚类的分片ICP配准矫正算法。较之整体配准下的矫正结果,本文所提算法在有效抑制偏离比(RP)"跑偏"现象的同时场景平均失配度(η)也降低了至少60%。
[Abstract]:As the key technology of video surveillance and 3D map construction, moving target detection and scene reconstruction are the foundation of real-time navigation obstacle avoidance and path planning. As a sub-problem of environmental perception task, they are closely related to the development of robot, unmanned aerial vehicle, unmanned vehicle, body sense game and so on, and are part of human intelligent life. The current research on this problem is based on a single vision or laser sensor, which is limited by its vision range, data volume, data richness, real-time and anti-jamming, etc. It is often difficult to meet the requirements of real-time multi-task scene. Based on the complementary and constraint of laser information and visual information, this paper designs a perceptual system which can simultaneously detect moving targets and reconstruct scene. The two sub-problems mentioned above are deeply studied in this paper. The main contents are as follows: 1. The differences between laser and vision sensors in the multi-sensor hardware platform Multisense-SL are compared and analyzed, and the rationality of the fusion of the two in dynamic scene perception is demonstrated. And based on the ROS software platform message publishing and subscription mode, A fusion framework. 2, which can simultaneously perform multiple real-time sensing tasks, is designed to solve the problem of "background exposure region" misjudgment and incompleteness of 3D foreground extraction in moving target detection based on visual background subtraction. A fusion moving target detection algorithm based on feature layer is proposed. Taking the laser foreground points based on the visual sorting as the heuristic information, the foreground clustering of 3D distance constraints under the neighborhood search of two-dimensional images is carried out, and the foreground map of two-dimensional images and the 3D scene cloud are obtained at the same time. It can effectively restrain the misjudgment of "background exposure area" while extracting foreground objects completely. Aiming at the noise problem of laser and visual data bottom fusion and the point cloud loss of registration directly through the external calibration relationship between sensors, it can be used to solve the problem of noise problem in the bottom layer of laser and visual data. The matching phenomenon is optimized in the fusion reconstruction of the scene. In the preprocessing stage, the field of view constraints, time constraints and occlusion constraints are proposed to effectively filter out most of the clutter points, and in the fusion optimization phase, a segmented ICP registration correction algorithm based on clustering is proposed. Compared with the correction results under global registration, the proposed algorithm can effectively reduce the deviation of deviation ratio (RP) and reduce the average mismatch (畏) of the scene by at least 60%.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】