面向智能导引的非结构环境主动视觉检测方法及验证系统

Active monocular vision such as PTZ camera，which can manipulate the viewpoint and focal distance of the camera in order to investigate the environment and get better information from it，has a broad prospect in the application of unstructured environment visual guidance for the mobile intelligent agent. At present, its technical bottlenecks are mainly in the following aspects: Distinctive-vision-feature based detection and recognition of a 3 dimensional object in real time；Real time accurate relative position estimation and measurement based on active monocular vision. Trying to find precept dealing with the issues，this project analyzes the situation and origin，focuses on the researches as follows: A novel automatic extracting vision feature method based on convolution network (CNN) deep learning structure, which combines the pool sampling and deep convolutional network with random forest classifier, is designed to get effective feature representation with better distinctiveness, which is called as adaptive pooling dense SCARF(Shape-color Alliance Robust Feature) hierarchical vision feature; Research and develop a fast target detection and recognition strategy based on the visual features in the image video; Search for both global and local invariant feature points as the image matching of the corresponding homonymous points. Then, by calculating the infinite homography matrix between two images of the counterpart points，a method of real-time accurate calibration of camera intrinsic parameters in the fixed focus mode and the strategy of focal length setting in zoom mode are put forward. The Longuet-Higgins binocular model is established to realize the on-line relative pose estimation with monocular vision. Finally, the feasibility and practicability of the proposed 3 dimensional target intelligent detecting and identifying active monocular vision module will be verified by constructing principle prototype. This active monocular vision module can be widely used for the mobile robots in the field of housekeeping, medical, aerospace and etc.

主动单目视觉因可控制目标物体成像大小和在视场中位置，在移动智能体的非结构环境视觉导引应用中前景广阔，目前其主要技术瓶颈是：①基于物体视觉特征的三维实物目标实时检测和识别；②基于主动单目视觉的实时准确相对位置估测。本项目将深入分析探讨技术瓶颈的成因及解决途径，拟设计基于深度学习的卷积神经网络，通过将池化采样、深层卷积网络与随机森林分类器相结合，以期提取出的新型池化密集可调分层视觉特征具有更强的可区别性；研究制定在图像视频中针对该视觉特征的快速目标检测识别策略；寻求兼具全局和局部不变性的特征点作为图像匹配的同名点；提出定焦模式下通过计算分解两幅图像的同名点之间的单应性矩阵而实时精准地标定出摄像机内参的方法以及变焦模式下标定出摄像机内参的策略，建立Longuet-Higgins双目模型以实现单目视觉在线相对位姿估测；构建原理性样机，验证所提出的三维目标智能检测主动单目视觉模块的可行性和实用性。

实例目标检测和三维位姿测量是计算机视觉领域的研究热点，在智能监控，视觉导引，人机交互、智能服务等领域有着重要的应用。本项目为了充分挖掘可摇摆变焦的主动单目视觉PTZ（Pan-Tilt-Zoom）相机在目标检测和位姿测量中的应用潜力，深入分析了目前在非结构环境实现三维实例目标的实时检测、识别和准确相对位姿估测所存在的主要技术瓶颈成因，完成了通用式平行双目视觉系统数学模型的构建，并基于单个PTZ相机和精密滑台创建了单双目智能视觉模块，该模块利用本项目的系列研究成果，可同时满足在线单目进行目标识别且双目精准测量该目标相对位姿的要求，实现了集搜索，检测，分割、测量于一体的全自动视觉导引功能。.为了在非结构环境中精准、快速、智能地检测实例三维物体目标，设计构建了基于人工提取特征的级联式分类器和可自动提取特征的卷积神经网络算法，重点研究了如何解决卷积神经网络在实例物体检测应用中通常需要千、万以上数量级的带标签训练数据进行有监督训练，而人工采集与标注工作既耗时又耗力的难点问题。研究成果包括: 1)设计并实现了由生成式反卷积网络GDDNE自动生成不同视角下的带标注实例物体样本和可提升鲁棒性的对抗式网络AGR-FCN所组成的双框架深度学习算法；2）构建了SVD RCNN和Gram-SS两种半监督神经网络，基于独创的关键样本选择与多样性保留的协同训练策略，仅需人工标注5%数量的样本就实现了基于少量样本和大量无标注样本而快速精确地检测实例物体。各算法的平均精确率皆高于80%。.为保证单双目智能视觉模块实现精准测量，针对现有特征提取及匹配算法的不足，创建了可有效提高彩色有畸变图像特征点匹配精度的CSCD-SURF描述子；提出了一种精度与张正友方法相当，无需靶标，对PTZ摄像机含畸变系数在内的5个内参进行实时高精度全自动在线标定的方法，并通过语义分割实现了对目标特征点的自动筛选与匹配，进而实现了一站式全自动目标位姿测量。测量范围在5米以内时位置的相对误差不超过0.4%，姿态角的绝对误差不超过1.5°。