惯性与高阶特征辅助的图像动态环境感知方法研究

Environmental perception is a critical issue of autonomous navigation for robots. Image-based environmental perception has been widely used in self-localization and accurate control in robotic systems as cameras are light weight and low power, while containing rich image information compared with other sensors. But existing image-based methods of environmental perception has two limitations: 1) They use the fixed feature points for location and map construction. If there are moving objects in the scene, those static feature points will be occluded, which leads to the degradation of stability and accuracy; 2) Without external aid, image-based environmental perception has scale ambiguity problem that become more difficult to be solved when fixed feature points are occluded...We propose to improve the current methods for image-based environmental perception to make them applicable to dynamic scenes. We plan to develop a new mathematical model that works for dynamic sense, and use the information from IMU sensors and geometric features in the scene to aid environmental perception – not only estimate ego motion (Localization), build scene map (Mapping), but also track moving feature points (Tracking) in a fast, accurate and stable way. We hope that our research will improve the capability of automatic navigation of robotic systems, especially for flying vehicles. .

图像环境感知方法在机器人自主定位导航与精确制导控制中被广泛应用。但目前图像环境感知方法尚存在如下局限：1) 使用场景中的固定特征点进行定位与地图构建，当特征点被常规场景中的运动物体遮挡时，机器人姿态估计与场景测量精度及其稳定性将大为下降；2) 若无其它辅助，纯图像环境感知方法存在尺度模糊性问题，在固定特征点被遮挡的情况下还将导致尺度浮动变化，导致环境感知性能恶化。.本课题提出结合惯性姿态测量信息和场景高阶几何信息来实现动态场景的图像环境感知方法。通过对目前基于静态场景假设的扩展卡尔曼滤波器扩展，利用惯性姿态信息对自身运动进行预测，并对固定特征点进行识别，结合高阶几何特征的观测模型进行状态求解，从而达到快速、稳定、精确地估算机器人自身姿态位置、场景三维地图结构，以及移动特征点三维运动轨迹的目的。期望通过本课题研究提高机器人在复杂场景下的自主导航能力。

本课题围绕着如何提高基于图像的视觉定位与环境感知方法的稳定性和精度进行研究，取得以下三个研究成果。第一，提出一种基于人造环境中普遍存在的结构线条的视觉SLAM方法-StructSLAM。 该方法减少了仅依赖特征点视觉SLAM的累计误差，并提升在纹理不佳情况下特征检测与跟踪能力。在机器人SLAM公用测试数据集rawseeds在同类方法取得性能第一的定位精度。第二，提出一种快速稳定的线条描述子。其工作原理与特征点描述子BRIEF类似，采取一种随机抽样的方式，经过离线训练，得出LBD线条描述子。实验结果表明，该描述子同时具有高性能和高效率的特点，非常适合视觉SLAM实际应用。第三，基于空气动力学的无人机自主定位方法 - 针对室内GPS缺失并且光照条件极端恶劣情况下视觉定位方法极易失效的问题，提出结合旋翼无人机特殊的空气动力学特性，与其他传感器融合, 实时估算无人机飞行速度和姿态。该方法在光照条件极端不好的情况下，仍然能准确地估算出无人机的速度，解决了纯视觉方法的短板。