面向移动机器人的目标检测与跟踪方法研究

CASIA OpenIR > 毕业生 > 硕士学位论文

	面向移动机器人的目标检测与跟踪方法研究
	张磊杰
	2017-05
学位类型	工学硕士
中文摘要	环境感知是移动机器人具备人工智能的重要且首要条件。场景三维信息获取、目标检测与跟踪等是环境感知的重要组成部分，在军事、安保、助老助残等方面具有广泛的应用前景。本文针对移动机器人目标检测与跟踪开展研究，论文的主要内容如下：首先，介绍了移动机器人及环境感知的研究背景和研究意义，阐述了立体视觉和激光雷达的研究现状，对目标检测和目标跟踪进行了综述，并对论文内容和结构做了介绍。其次，开展了场景三维信息的获取研究。面向双目视觉，完成了内外参数的标定以及立体匹配，获得了场景的深度图；面向单目与多线激光雷达，实现了单目与多线激光雷达的联合标定，由此将激光点云投影到图像坐标系，从而提供了图像目标区域的三维信息。第三，研究并改进了基于深度学习的目标检测方法。在提升实时性方面，通过裁剪SSD目标检测模型，训练得到可满足嵌入式GPU平台实时性要求的目标检测模型。在提高准确性方面，结合场景三维信息，对基于深度学习的目标检测提供的候选定位框作进一步筛选与精定位以实现目标跟踪的初始化。第四，结合KCF跟踪器，开展了目标跟踪方法的研究。基于单目视觉的目标跟踪方面，面向目标跟踪中存在的模型漂移、目标部分遮挡等问题，设计级联定位器以实现裁剪后SSD检测模型与KCF跟踪器的有机结合。基于双目视觉的目标跟踪方面，利用场景三维信息提取目标的深度分布，并设计了融合HOG、CN、LDP的多特征KCF跟踪器，用于应对目标全遮挡、模型漂移等问题。第五，搭建了移动机器人目标检测与跟踪实验系统，描述了移动机器人硬件平台及其软件框架。对于装有双目视觉传感器的移动机器人，采用基于场景三维信息的多特征KCF目标跟踪方法；当机器人通过单目视觉和多线激光雷达感知环境时，它采用综合HOG、CN的KCF目标跟踪方法，并通过多线激光雷达获得目标的三维信息。室内/室外环境下移动机器人跟随行人的实验验证了上述方法的有效性。最后，对本文工作进行了总结，并指出了需要进一步开展的研究工作。
英文摘要	Environmental perception is essential for the robot with artificial intelligence, and it should be firstly resolved. As the important parts of environmental perception, three-dimensional information acquisition as well as target detection and tracking has a wide potential applications in military, security and elder and disabled people assistance. This thesis focuses on the research on target detection and tracking for mobile robot. The main contents are as follows: Firstly, the research background and significance of mobile robot and environmental perception are introduced. The research development of stereo vision and lidar is presented. The target detection and target tracking are then reviewed. The contents and structure of the thesis are also introduced. Secondly, the research of 3D information acquisition is conducted. For binocular camera, the calibration of internal and external parameters and stereo matching are completed, and the depth map is acuqired. For monocular camera and multi-channel lidar, on the basis of the joint calibration, lidar point cloud can be projected into image coordinate system, which provides 3D information related to the target. Thirdly, the method of target detection based on deep learning is studied and improved. To improve the real-time performance, cropping is applied to the SSD detection model, which can meet the real-time requirement of the embedded GPU platform. In the aspect of accuracy improvement, the result of target detection based on deep learning can be further precisely located with the combination of 3D information, which is used to achieve the initialization for subsequent target tracking. Fourthly, the target tracking approaches based on the KCF tracker are presented. For monocular target tracking, a cascade locator is designed to realize the combination of the cropped SSD detection model and KCF tracker, which can alleviate the model drift and target partial occlusion problems. As for binocular target tracking, based on the depth distribution of the target extracted from the 3D information of the scene, a multi-feature KCF tracker with HOG, CN, LDP is designed to deal with the problems of occlusion and model drift. Fifthly, a robotic experimental system for target detection and tracking is built. The mobile robot platform with its software framework is described. The mobile robot with binocular camera adopts the multi-feature KCF tracking approach based on the 3D information of the scene. When the robot senses the environment using monocular camera and multi-channel lidar, it adopts the KCF approach with HOG and CN for target tracking, where the target-related 3D information is provided by the lidar. The indoor and outdoor experiments of pedestrian following conducted on the mobile robot platform verify the effectiveness of the afore-mentioned approaches. Finally, the conclusions are given and future work is addressed.
关键词	双目视觉激光雷达深度学习 Kcf 目标检测与跟踪移动机器人
文献类型	学位论文
条目标识符	http://ir.ia.ac.cn/handle/173211/14800
专题	毕业生_硕士学位论文
作者单位	自动化研究所
推荐引用方式 GB/T 7714	张磊杰. 面向移动机器人的目标检测与跟踪方法研究[D]. 北京. 中国科学院研究生院,2017.

条目包含的文件
文件名称/大小	文献类型	版本类型	开放类型	使用许可
zlj_移动机器人目标检测与跟踪研究_5（5312KB）	学位论文		限制开放	CC BY-NC-SA