基于RGBD图像的移动机器人避障策略研究

发布时间：2018-07-24 18:41

【摘要】：随着经济的发展,城市人口的逐渐增多,城市中心医院的护士在临床护理工作也越来越繁重,护患矛盾时有发生。为了将护士从平时繁琐的工作中解放出来,更好地为患者服务,将护士助手移动机器人引入医院中,代替护士完成输送问询等工作。医院环境中人流量大,要求护士助手移动机器人在其中安全可靠地完成物品运输工作极负挑战性。为此,本文针对室内动态环境下的移动机器人行人目标检测跟踪与避障策略问题展开研究,基于ROS系统进行高效地开发与验证。提出的避障策略流程为:基于人体识别算法对机器人前方正在行进的人体进行检测确定其所在的图像区域,对确定的图像区域进行目标跟踪,通过这种方法我们可以获得前后两个时间段的人体相对于机器人的位置,进而确定人体目标的运动状态,经过IMM滤波后获取人体的准确运动状态,最后基于改进的人工势场法进行局部路径规划完成避障行为。具体工作如下:与单独使用RGB图像来进行检测不同,基于信息更加丰富的RGB-D图像进行行人检测。提出的算法流程为:首先基于密度统计信息对点云图像进行鲁棒高效地聚类分析获得各点的集合,然后根据高度标准划分出人体存在的可能区域,最后基于SVM-HOG框架进行人体检测。实验表明,采用上述方案在降低误检率的同时也可以降低计算机的运算量从而达到实时性的要求。采用Online Boosting算法将目标跟踪问题转化为一种二值分类问题进行高效处理。该算法通过对人体所在区域进行Haar-like特征提取,实时训练更新所需要的分类器。当下一次检测样本到来时,使用训练好的分类器对该样本进行判别并重复上一步骤实时更新,以此往复。另外还引入了贪婪数据关联逻辑方法以解决当目标丢失后如何重新进行跟踪问题。基于交互多模型(IMM)滤波算法准确估计障碍物的运动状态,采用改进的人工势场法进行局部路径动态规划。针对卡尔曼滤波器在实验机器人上估计障碍物位置精度差的问题,提出基于交互多模型的滤波算法。对传统的人工势场法进行改进,将障碍物的速度纳入考虑范围内,相较于传统方法,本文提出的算法使机器人在动态环境下能更好的完成避障,减少与障碍物的碰撞率。综合实验表明,基于上述策略设计的护士助手机器人在传感器检测范围内,能够处理稀疏人群正常步行速度条件下的室内动态避障。与传统避障策略相比较,在机器人在规划效率有所提升,对行人的处理上更加智能,更加安全。
[Abstract]:With the development of economy and the increase of urban population, the nurses in urban central hospital have more and more heavy clinical nursing work, and the contradiction between nurse and patient occurs from time to time. In order to liberate nurses from the usual tedious work and better serve the patients, the nurse assistant mobile robot was introduced into the hospital, instead of nurses to complete the transportation of inquiries and other work. In the hospital environment, the mobile robot is required to carry out the transportation of goods safely and reliably. Therefore, this paper studies the problem of pedestrian target detection, tracking and obstacle avoidance strategy of mobile robot in indoor dynamic environment, and develops and verifies it efficiently based on ROS system. The flow chart of obstacle avoidance strategy is as follows: based on the human body recognition algorithm, the moving human body in front of the robot is detected to determine the image region, and the target tracking is carried out to the determined image region. Through this method, we can obtain the position of the human body relative to the robot in the two time periods before and after, and then determine the moving state of the human body target, and obtain the accurate motion state of the human body after IMM filtering. Finally, local path planning based on the improved artificial potential field method is used to accomplish obstacle avoidance. The main work is as follows: different from using RGB images alone, pedestrian detection is based on more informative RGB-D images. The proposed algorithm flow is as follows: firstly, based on the density statistics, the point cloud images are clustered to obtain the set of points, and then the possible regions of the human body are divided according to the height criteria. Finally, human body detection is carried out based on SVM-HOG framework. Experimental results show that the proposed scheme can not only reduce the false detection rate but also reduce the computational complexity of the computer so as to meet the real-time requirements. Online Boosting algorithm is used to transform the target tracking problem into a binary classification problem. By extracting the Haar-like features of the human body, the algorithm can train the classifier needed for updating in real time. When the next detection sample comes, the trained classifier is used to distinguish the sample and repeat the previous step to update the sample in real time. In addition, the greedy data association logic method is introduced to solve the problem of how to track the target again when the target is lost. Based on the interactive multi-model (IMM) filtering algorithm, the moving state of obstacles is estimated accurately, and the improved artificial potential field method is used for local path dynamic planning. Aiming at the problem of poor accuracy of Kalman filter in estimating obstacle position on experimental robot, a filtering algorithm based on interactive multi-model is proposed. The traditional artificial potential field method is improved and the velocity of obstacles is taken into account. Compared with the traditional method, the algorithm proposed in this paper can make the robot avoid obstacles better and reduce the collision rate with obstacles in dynamic environment. The comprehensive experiments show that the nurse assistant robot designed based on the above strategy can deal with the indoor dynamic obstacle avoidance under the condition of the normal walking speed of the sparse crowd within the detection range of the sensor. Compared with the traditional obstacle avoidance strategy, the robot is more intelligent and safer in planning efficiency and pedestrian handling.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP242

【参考文献】