中国科学院自动化研究所机构知识库

感知和持续观测运动目标是机器人克服动态环境挑战必不可缺的能力。要实现机器人对运动目标的感知和持续观测需要对两方面的技术进行研究：一是非固定的相机在场景中检测运动目标；二是根据目标在场景中的变化情况，实现移动机器人的运动控制设计，通过机器人对运动目标的稳定平滑跟随，将其保持在视野中合适的位置上。光流是运动目标与相机之间的相对运动在相机图像平面上的投影，一定程度上反映了目标的运动情况，对运动目标的检测、跟随控制具有重要的指导意义。本文基于光流对运动目标检测以及机器人跟随目标过程中的运动控制两项关键技术展开研究。主要的工作和贡献有：

（1）提出了一种基于光流的实时运动前景分割算法。现有的实时运动前景分割法一般基于背景减除法设计，背景模型容易受到前景污染，同时容易受光照变化、目标纹理简单等因素影响，检测效果难以满足后续运动目标检测任务的需求。为了提高实时算法的性能，本文提出使用光流作为运动前景和背景区分的主要依据。通过在前景背景混合的光流场中回归背景光流分布作为参考，结合双判断机制和自适应的判别阈值，鲁棒地提取运动前景。在这过程中使用了随机采样一致算法和自适应的光流估计区间，保证回归的精度。实验证明本文方法不但具备良好的实时性，而且在公开数据集DAVIS2016上MJ综合指标达到0.536，相比已有的实时算法MCD5.8ms和SCBU分别有0.242和0.356的提升。

（2）提出了一种基于快速聚类分析和微调参考框的运动目标检测算法，并构建了运动目标检测算法测评系统用于算法性能评测。运动目标检测面临着目标类别不确定、数量不确定、形状不规则、相互遮挡和存在显著阴影等诸多挑战。为了能应对上述挑战，本文利用快速聚类分析对场景中的运动目标进行数量分析和初步定位。然后使用微调参考框的方法进行边界框定位，以应对冗余聚类中心和目标形状不规则等问题，减少假阳边界框并提高边界框定位的精度。针对已有工作所使用的数据集存在属于私有、数据量不足，指标不统一等问题，本文基于已标注好的公开数据集构建运动目标检测算法评测系统，选择用于评测算法的指标，并根据指标和数据集开发了对应的开源测评工具。

（3）提出了一种基于光流的运动目标跟随伺服控制方法。实际应用中运动目标的复杂运动规律要求控制器在目标速度变化时，能作出快速的响应以及有效的调整。针对该问题，本文通过推导给出了光流与目标运动速度之间的数学表达关系，实现了目标运动速度估计器，并将其与基于图像的视觉伺服控制技术相结合，提出了基于光流的增量式人体目标跟随控制器。该控制器不但能够实现基于图像特征偏差的运动调整，而且能够依据光流及其偏微分信息实现面向跟随的目标速度补偿。综合利用仿真和实际实验对所提控制方法进行了验证，证明本方法在应对目标速度变化时，同时具备响应快、超调小和误差收敛迅速的特点。

Challenges in dynamic environment make a request upon robots for detecting and following-up the moving objects. To solve this problem, two aspects of technology are needed to be studied: One is a detector for efficiently detecting moving objects in the scenes with an Non-stationary camera, and the other is a visual servo controller for robot stably following the detected moving objects. This paper leverages optical flow as the main cue for researching moving object detecting and following, as it is the projection of the moving object's relative motion onto the image plane, and to some extend, showing some features of the object motion. The main work and contributions are as follows:

(1) A method of real-time moving foreground segmentation based on optical flow is proposed. Most existing real-time moving foreground segmentation methods follow the idea of modeling based background subtraction, and thus, perform unsatisfactorily when used with a moving camera. The proposed method leverages optical flow as the main cue to solve this problem and produces high quality segmentation result while maintains the real-time property. It achieves this by two key process: one is using second-order polynomial and RANSAC algorithm to accurately regress the distribution of the background optical flow for reference; and the other is using a dual judgment mechanism with an adaptive threshold to robustly identify the moving foreground. The proposed method scores 0.536 on MJ metric upon DAVIS2016 dataset, making progress by 0.242 and 0.356 respectively when compared with the existing real-time methods MCD5.8ms and SCBU. 

(2)A method of moving object detection is proposed. The proposed method based on moving foreground segmentation so that it can dual with all kinds of moving object. Besides, it takes advantage of fast clustering technologies to fleetly analyze the composition of the moving foreground and initially locate each instance. The bounding box results is generated by selecting and fine tuning some referenced boxes which can do help to precisely locate the bounding box and reduce false positive results. We also develop a relatively more effective evaluation system for moving object detecting algorithm performance analysis. With this evaluation system, we show the feasibility of the proposed method.
%which is according to the variance of the detected object in the scenes.

(3)An optical flow based object following controller is proposed. In practical applications, the unclear object motion makes request on the following controller for quick response and effective adjustment in times of the change of object motion. The proposed following controller estimates the relative velocity of the object based on the object's optical flow, and takes both object velocity and positional error into consideration when modifies its output. In this way not only can it adjust the motion of the robot according to the deviation between the reference features and the actual features, but also can compensate the object velocity effectiveness. In simulation experiments and practical experiments, we show that the proposed following controller is equipped with the advantages of quick response, small overshoot and high accuracy.