Design And Implementation Of Interactive Navigation System For Mobile Robot In Dynamic Environments

With the development of robotics,navigation in dynamic and complex unstructured environments has increasingly become a research hotspot for mobile robots.Traditional navigation methods have problems such as poor path efficiency,low navigation success rate,and poor pedestrian comfort experience in complex environments.To this end,according to the dynamic and complex working environment of indoor mobile robots,this thesis considers pedestrian movement trends and the interactivity of objects,and constructs a semantic map containing pedestrian information and interactive object information by introducing pedestrian perception and interactive object recognition.A deep reinforcement learning interactive navigation method combined with semantic maps is studied.For the problem of pedestrian perception and modeling of mobile robots in dynamic environments,this thesis uses a laser sensor to recognize human legs based on the Adaboost algorithm and a RGB-D camera to recognize the upper body based on deep template matching,and realizes multi-modal pedestrian detection through multi-sensor information fusion.On this basis,the trajectory motion model is constructed and the pedestrian trajectory tracking and management are realized based on the data association method.Furthermore,in view of the problem that traditional trajectory prediction methods cannot fully extract pedestrian interaction information,a generative adversarial trajectory prediction method based on attention mechanism is proposed.The attention pool layer is constructed to extract the influence of surrounding pedestrians on the target pedestrian's future trajectory,and experiments on the ETH and UCY datasets have proved the accuracy of the proposed method.For the problem that traditional maps do not contain interactive semantic information and cannot meet the interactive navigation of mobile robots,this thesis uses the YOLO algorithm to identify interactive objects,performs semantic segmentation based on the recognition results,and uses overlay grid mapping algorithm to construct interactive semantic local grid map for local path planning of mobile robots.On this basis,the interactive layer costmap is built to integrate the information of interactive objects,and the social layer costmap is constructed based on the pedestrian trajectory prediction results,and the individual comfort of pedestrians is modeled by the method of establishing Gaussian distribution cost.Finally,they are integrated into the global map and used for global planning of mobile robots to improve pedestrian comfort experience and navigation efficiency.For the problem that traditional path planning methods cannot plan effective paths in complex environments,and most existing reinforcement learning navigation algorithms based on mapless vision have poor generalization ability in real scenes.In this thesis,through Markov decision process modeling,a PER-DQN deep reinforcement learning local path planning algorithm combined with semantic maps is proposed.By using continuous multi-frame semantic maps and the relative pose of the robot and the target point as the network input,the training optimization strategy of multi-stage curriculum learning is used to train the robot in simulation environments.The comparison experiment verifies that the proposed method can converge faster and the generalization performance is better.On the basis of the above-mentioned technical research,an interactive navigation system for mobile robots has been constructed,and functional modules such as pedestrian perception and trajectory prediction,semantic map construction and path planning have been developed.Through the establishment of simulation environments and real environments,the overall experiment and comparative test of the mobile robot interactive navigation system have been carried out to verify the effectiveness and reliability of the overall system architecture and various functional modules.