室内单目机器人视觉目标发现与跟随

摘要

本文研究了室内单目机器人上的视觉目标人发现与跟随问题,分为场景变化检测,目标人检测,目标人视觉追踪和目标人主动跟随几个部分,主要研究了场景变化检测算法和目标人视觉追踪算法.高速的场景变化检测算法通过对场景建模来分析该场景是否变化,为目标检测部分提供潜在变化帧和潜在变化区域.实验结果表明能够提高系统运行速度,减少机器人运行时的卡顿.视觉目标追踪算法结合表观模型和SLAM过程得到的地图点信息,估计目标区域内属于背景的部分,减少由于遮挡和目标尺度变化对于追踪算法的表观模型的影响,实验结果相比于对比算法取得较大效果提升.本文尝试使用近年来效果较好的深度神经网络来进行目标检测.使用小型深度网络并加强对于室内场景下人的学习,在运行速度和检测效果方面取得较好的平衡.在视觉目标人的发现和跟踪的基础上,我们实现了机器人的跟随.由于单目视觉仅能够提供目标的方向信息,所以机器人主动跟随的目标是保持目标人在成像平面的水平居中位置.在目标无遮挡和部分遮挡的情况下,机器人能够成功的跟随人.%This study researches on the visual detection and following of object people by indoor monocular robots, which includes scene change detection algorithm, visual object people detection algorithm, visual object tracking algorithm, and robot following, with focuses on scene change detection algorithm and visual object tracking algorithm. A high speed scene change detection algorithm judges whether the scene changes by constructing scene models. If the scene changes, the algorithm outputs the change region, which is used by the visual object detection algorithm. The experiment shows this algorithm speeds up the system and alleviates the latency of robots. The visual object tracking algorithm combines the appearance model and map information obtained in SLAM process. The map information can judge which part of object bounding box is actually the background, which can reduce the effect of occlusion and object scale change on appearance model. This algorithm improves visual object tracking performance in the experiments. This paper applies the latest deep neural networks to do visual object people detection. We train a small deep neural network with enhancement on indoor people, which achieves a good balance between running speed and detection performance. Based on the visual detection and visual tracking of target, we accomplish robot following. Since monocular robots can only get the bearing information of target, the goal of robot following is keeping the target in the horizontally middle point of image plane. The robot can successfully follow human even if the person is partially occluded.