Robot Navigation In Complex And Dynamic Pedestrian Scenarios

Navigation is a very important and fundamental ability for mobile robots.With the development of artificial intelligence,robotics,and sensor technology,more and more mobile robots are applied to open and crowd environments.In order to cope with such dynamic and complex scenarios,robots are equipped with more and more sensors to per-form complex perception tasks.The accurate calibration of the sensors is a prerequisite for the successful navigation of the robot.The general batch calibration framework based on motion capture system can solve the problem of multi-sensor calibration,but the current system is only limited to laser and robot calibration,and the accuracy of the core extrinsic calibration method is not high.In this dissertation,we propose a reliable and accurate calibration method of RGB-D cameras based on the general batch calibra-tion framework and motion capture system(MoCap),which extends the general batch calibration framework to the field of color or depth camera calibration.In the extrinsic calibration,we make full use of the characteristics of the MoCap system to customize the global refinement step for the hand-eye calibration,and it further improves the accu-racy of the extrinsic calibration.Moreover,a non-recursive and novel data acquirement method is used to get the ground truth of every pixel in depth images and a one-step,model-free depth correction approach is applied to obtain the parameters of the depth correction models.Compared with the existing depth correction techniques,our method can simultaneously estimate the mean and variance of the depth error at different mea-surement distances.The experimental results show that our calibration method greatly improves the matching and measurement accuracy of color point cloud.Autonomous navigation of mobile robots relies on safe and efficient obstacle avoidance algorithms.In recent years,a large amount of research work has begun to ex-plore the application of deep reinforcement learning(DRL)for robot navigation in the dynamic environments,and end-to-end navigation approaches based on raw sensor data or pedestrian position information has been developed.In this dissertation,we propose a heterogeneous multi-robot obstacle avoidance algorithm based on grid map and DRL in a communication-free environment.We use the egocentric local grid map to repre-sent the environmental information including robot's shape and observable appearances of other robots and obstacles,which can be easily generated by using multiple sensors or sensor fusion.Compared to other methods,the map-based approach is more robust to sensor noise,does not require robots' movement data and more efficient and easier to be deployed to real robots.We first train the neural network in a specified simulator of multiple mobile robots using DPPO,where a multi-stage curriculum learning strat-egy for multiple scenarios is used to improve the performance.Then we deploy the trained model to real robots to perform collision avoidance in their navigation without tedious parameter tuning.We evaluate the approach with multiple scenarios both in the simulator and the real world.Both qualitative and quantitative experiments show that our approach is efficient and outperforms existing DRL-based approaches in many indicators.However,if the obstacle avoidance algorithm treats pedestrians as ordinary obsta-cles or robots,it may lead to close obstacle avoidance effects and bring discomfort to pedestrians.At the same time,the walking strategy of pedestrians is generally different from the obstacle avoidance strategy of robots.The crowd navigation does not com-pletely meet the applicable conditions of the multi-robot obstacle avoidance algorithm.Therefore,we expand our map-based multi-robot obstacle avoidance approach,add the pedestrian map channel generated by the multi-sensor pedestrian perception module,and design a new reward function part and the multi-strategy pedestrian training en-vironments.Experiments show that our method further improves the success rate of obstacle avoidance in dynamic pedestrian environments.At the same time,robots will also encounter more serious localization failure problems in a dynamic pedestrian en-vironment.Therefore,we propose a dynamic localization method based on pedestrian perception filtering to improve the accuracy of robot localization.However,the passive localization method still cannot completely solve the problem of sensors obscured by a large number of dynamic human legs,so we propose a joint active relocation method based on QR code and laser sensor.When the localization error is large,the robot moves to the QR code recovery area and calibrates the global pose autonomously.Ex-perimental results show that our relocation method can relocate the robot's global pose very accurately even when the localization is completely lost.Finally,based on the above research results and "criteria of closeness",we propose a specific deployment plan that attempts to solve the long-term autonomous navigation in open and crowd en-vironments,and successfully conducted a large-scale long-term navigation test at the Chengdu Research Base of Giant Panda Breeding.