Study On All-day SLAM Algorithm For Field Robots

Simultaneous Localization and Mapping(SLAM)is of great significance to outdoor robots and autonomous driving,whose rapid development puts forward higher requirements for the performance in complex environments.In the field,it is necessary for SLAM system to overcome the problems such as high dynamic illumination,low visibility at night,and cross-time localization of the environment in all-day.Traditional visual-based SLAM(VSLAM)or multi-sensor fusion SLAM is very easy to fail under above-mentioned circumstances.In recent years,compared with visible cameras,thermal infrared cameras have better imaging quality at night and other dynamic conditions such as haze,and have the advantage of working in all-day.Thermal infrared cameras have gradually been applied in the SLAM field,however,weak texture and low signal-to-noise ratio limit the development.The existing infrared SLAM solutions are mostly based on traditional VSLAM frameworks,whose accuracy needs to be further improved.Therefore,this thesis is committed to develop an infrared visual odometry that is more suitable for infrared cameras,more research revolves around infrared-lidar SLAM framework to overcome the above-mentioned challenges.In order to realize the autonomy of outdoor robots in all-day,and further improve the stability in complex environments,and achieve efficient real-time localization,this thesis proposes an infrared visual odometry method based on 3D-2D edge ICP matching.Combining the advantages of the traditional feature-based method and the direct method to perform direct registration of image edge features,this method avoids the defects of infrared camera such as texture-less and high temperature drift,that it is more suitable for infrared images.There also develops a tightly coupled infrared-lidar SLAM framework to fuse the infrared camera and lidar information for robust localization.By using threedimensional lidar pointclouds to construct a dense depth map to assist visual feature restoration of scale information,infrared visual odometry is used to perform pointclouds nonlinear motion model de-distortion.At the meantime,for scene recognition in the allday environment and successful relocalization in the maps constructed at different time periods,this thesis proposes a scene recognition verification method based on infrared images,which improves the success rate and stability of the scene recognition module in the all-day environment.By using the loop constraint,we could adjust the global pose,and obtain a more accurate environment map and localization results.In this thesis,the algorithm is verified under the self-made campus environment datasets.The results show that the infrared-lidar SLAM proposed in this thesis could achieve stable real-time localization and scene recognition in different periods in all-day environment.