Research And Implementation Of Robot Navigation Method Based On Multi-view Camera Perception

Robot navigation is a general and basic task in the field of robotics,and its technology can be applied to various fields such as autonomous driving,national defense,logistics,patrol and security check,etc.In recent years,the development of artificial intelligence technology,especially the development of deep learning and computer vision,has brought new opportunities for the development of research in robot navigation.Numerous researchers in industry and academia have carried out abundant research in this field.Although there has been a lot of classic work in this field,on the one hand,the complexity of system design,the demand for datasets and the interpretability of models have always been difficult to balance.On the other hand,most of the work uses single camera sensor for sense perception,which brings hidden dangers for safe navigation.In this paper,we propose to use multiple camera sensors for scene perception,but the simple feature fusion of multiple cameras may cause causal confusion of information,resulting in the model relying too much on the information of a certain view.Based on this,we propose to treat each input from an individual camera as a separate channel,and propose a series multi-task visual perception networks.These networks receive the image data from different cameras and is used to predict navigation-related tasks.For robot outdoor global navigation function,we use third-party maps API according to the GPS information of stating point,target point and robot localization for path planning on 2D map.According to the map with path planning information,we use neural network to predict the direction of the robot under a crossroads.Our overall algorithm combines a variety of technical framework,on the one hand,we use the end-to-end imitation learning to reduce the complexity of system design.On the other hand,according to the results of the network output,we propose a series of navigation indicator for environment low dimension representation.We use the rule-based method according to these indicators to design the controller for the robot navigation,which provide a more interpretable and safe control command.In this thesis,the neural network proposed above is experimentally analyzed on the validation set and the algorithm is applied to the real robot platform.We tested the navigation performance of the algorithm in a real scene and compared the performance of our algorithm with some classical algorithms.Finally,it is proved that our algorithm has good performance both in local obstacle avoidance and global navigation.