Research And Implementation Of All-weather Environmental:Perception System For Autonomous Mobile Robot

With the continuous development of deep learning technology,mobile robots have also been evolving to become more intelligent.The environmental perception ability of robots is the foundation for autonomous mobile robots to plan and make decisions,and Simultaneous Location And Mapping(SLAM)is the key technology for robots to realize environmental perception,which utilizes the data collected by sensors on the robots to perform the robot pose estimation and mapping.In order to improve the perception ability of robots in all-weather environment,a scheme combining camera and lidar was used in this thesis,and the adaptability of the camera to light changes was also enhanced by deep learning technology.The artificially designed image features used in traditional visual SLAM present quite different location of feature points extracted from the same object at different times and scales,making it difficult to apply in complex environments.However,depth features that are deeper image features extracted from a large amount of data showed stronger robustness and generalization ability than manually designed features.Therefore,in this thesis,an all-weather environmental perception system for robots based on Robot Operating System(ROS)was designed and developed,and cross-platform application framework Qt was also used for visual development.Through requirements analysis of functions,relevant functional modules and the overall framework of the system were designed.The main research contents are as follows:(1)HF-Net network was utilized in this thesis to extract image feature points for enhancing the adaptability of image features to external environmental changes such as illumination,season and scale.(2)A series of mismatching elimination strategies,such as bilateral nearest neighbor matching,distance threshold and Random Sample Consensus(RANSAC)algorithm were adopted in this thesis to ensure the accuracy of feature matching and thus improve the accuracy of visual pose calculation.(3)A method for depth estimation of image feature points was proposed in this thesis to estimate the depth of image feature points using the depth value information detected by lidar.(4)The methods mentioned above were incorporated into Visual-lidar Odometry and Mapping(VLOAM),a traditional visual laser fusion framework in this thesis to build the environmental perception system integrating HF-Net,the camera,and lidar.Finally,experiment and testing results show the solutions for the problems on the environmental perception of robots in this thesis are effective.