Research And Implementation Of Key Technologies Of A Robot Navigation System For Transport Applications In Indoor Complex And Dynamic Environments

Mobile robots play an important role in the intellectualization and informatization of logistics and transportation systems.Their application fields are gradually expanding from strictly controllable factory environments to more natural daily scenes.However,the complexity and dynamics of these scenarios have brought many challenges to the application of mobile robot technology.For example,the uncertainty and incomplete observability of environments increase the unreliability of the system,insufficient capacity of the robot perception system causes decision information deficiency;complex environments make it difficult to design the planning algorithm;how to deal with the interference of dynamic objects to the localization algorithm;the complex human-robot interaction problem and so forth.In this dissertation,we take robot navigation in complex and dynamic environments as background,and conduct research in three domains:laser-based pedestrian detection and tracking,robot localization,multi-level decision-making framework and obstacle avoidance.We resort to improving the performance,safety,and reliability of the robot navigation system by applying deep learning,sampling method,and dynamic programming.The main contributions are as follows:1.The problems of laser-based pedestrian detection and tracking in crowded environments are investigated.Firstly,to address the time-consuming and laborious problem of manually annotating laser data,we propose an automatic annotation algorithm.Secondly,the existing two-dimensional lidar datasets contain too few pedestrians.Therefore,we collected a large amount of laser data in crowded environments and automatically labeled it,which makes a contribution to open-source datasets.Finally,to address concerns that the inference speed of pedestrian detectors is slow on low-power platforms,we propose a preprocessing method based on jump distance segmentation.The experimental results show that the automatic annotation algorithm can quickly and accurately label the category of each laser point,and the jump distance down-sampling method can improve the reasoning speed by an order of magnitude on the premise of ensuring the performance of the model.2.The problems of robot localization and relocation in dynamic environments are studied.Firstly,to address concerns that the interference and damage of dynamic objects to the measured data,an outlier removal method based on the pedestrian detector is proposed to alleviate the influence of noise on the algorithm.Secondly,aiming at the problem that the location algorithm may fail in complex environments,we design a location anomaly detection algorithm.Finally,aiming at the problem that the Monte Carlo global localization method is inefficient and needs multi-step iteration to converge,a localization algorithm framework combining multi-resolution correlative scan matching and particle filter is proposed.Experimental results show that this method improves the efficiency and practicability of the location algorithm.3.The solution of robot navigation in complex and incompletely observable environments is studied.Firstly,in order to make the human cooperate with the robot better,we propose the navigation legibility index.Secondly,we decompose the navigation problem,and design suitable behaviors for the indoor transportation robot.Finally,this paper designs an obstacle avoidance and yield algorithm based on Frenet coordinate system.The experimental results show that the proposed algorithm improves the consistency and predictability of the robot’s behavior,and thus improves the efficiency and security of navigation.To sum up,the three main innovations of this dissertation are summarized as follows:1)We propose a simple and effective automatic pedestrian annotation algorithm based on two-dimensional lidar;2)We propose a robot localization framework based on correlative scan matching and particle filter;3)We propose an obstacle avoidance algorithm based on Frenet coordinate system,which considers the legibility,predictability and the right-hand traffic rule.Finally,we design and implement the software and hardware system of the indoor transportation robot prototype,and conduct a large number of tests in a real hospital environment(The First Affiliated Hospital of USTC)to verify the effectiveness of the algorithm,and analyze some unsolved challenges faced by the robot navigation system in complex and dynamic environments.