Study Of Farmland Perception System Based On Semantic SLAM

Agricultural production is gradually becoming intelligent and automatic.In the future automatic harvesting scenarios,the harvesting mechanism should be able to perform efficient and intelligent tasks planning based on the surrounding environment perception data.The tasks include global path planning,local dynamic path adjustment,and intelligent control of modules such as headers.In order to achieve these tasks,the key is the perceptual data.For example,the 3D map of farmland as the global perception data,can provide observers with long-term and traceable digital data for judging crop growth,and local crop distribution information can help the harvester decide the next step of motion.As a solution for constructing a 3D map of farmland,a real-time localization and mapping(SLAM)algorithm provides the basic technique.However,the traditional SLAM technology pays much less attention to the map construction than the localization.Purely dense point cloud maps can often only be used for appreciation and cannot be applied to the intelligent perception.To this end,some researchers have proposed semantic mapping in conjunction with environmental semantic information,which introduces semantic information to the prior identification and classification of data in the map.However,maps constructed by these methods often have bad noise,which is caused by motion blur from environmental dynamic objects.In addition,the dense maps generated are often too large and difficult for quick application.In view of the above problems,this paper proposes a semantic SLAM structure.In order to fuse semantic information and traditional SLAM algorithms,a multi-threaded fusion method is proposed.Semantic segmentation is used as an individual thread.Secondly,in view of the possible bad noise in the localization process of semantic SLAM,this paper proposes a method of bad dynamic object detection based on the combination of semantic segmentation and multi-view geometry.Through the prior dynamic semantic object mask and multi-view dynamic object mask based on optical flow tracking combined with epipolar search,the dynamic object region point set from tracking thread is eliminated to improve the pose estimation accuracy.Thirdly,the problem of how semantic SLAM introduces semantic information in the map construction process and reduces the difficulty of map reuse is addressed.This paper proposes to use the occupancy probability update method to assign grid-level semantic labels to octree maps.The allocation process retains repeatedly observed map grids through the "observation-update" strategy,and reduces accidentally observed noise.Finally,for how to obtain local crop distribution information in farmland,this paper proposes a multi-sensor-based crop distribution detection algorithm.In order to verify the function of the semantic SLAM system and the crop distribution detection algorithm,this paper conducted experiments on public data sets and actual farmland scenarios.Based on the public data set,the function of the semantic SLAM system to solve the bad noise in localization was tested.At the same time,the function of the semantic SLAM to map the semantics into octree was tested on both public data set and the farmland data.The experimental results show that the semantic SLAM system in this paper can complete the function of constructing 3D semantic maps in the farmland environment and cope with the presence of bad dynamic features,and the crop distribution detection algorithm can complete the accurate measurement of crop heights and boundary lines in the farmland.