Obstacle Detection Based On Point Cloud In The Application Of Agricultural Navigation

Various sensors are utilized to assure safety in the vicinity of a intelligent robot while it is walking,with the development and wide application of intelligent navigation.Among them,stereo cameras are prevalent these years due to its capability of distance measurement.According to the limitation of traditional detection methods in stereo vision,a method based on point clouds was presented to meet the demand of obstacle detection on robot while path planning.Point clouds of environment space with obstacle involved was taken as the object,a validity box was applied into space to eliminate point clouds in irrelevant regions.From top view,the area was divided into grids and the number of points in each grid was denoted as density of point clouds.Once the point cloud density in different ranges was calculated,a curve that fits the distribution of density,decreasing due to the range was figured out.It indicated that obstacles far from the camera created less point clouds than the near ones,leading to a miss match and recognition of obstacles.In order to compensate the sparse point clouds of obstacles in long range,density were compensated according to a curve of density-range relationship.And obstacles were able to be recognized by setting a threshold,so was the distance measured.The specific space that obstacle occupied was confirmed by setting one more validity box.Therefore its shape and size could be measured after projecting it onto front view.Experiments were conducted on the campus of Nanjing Agricultural University,Nanjing,Jiangsu Province,source codes were programmed and compiled by Matlab software.Experimental results showed that this method could restore obstacle information of point clouds effectively,while in the distance measurement test,it showed a maximum detection range of 28 meters and average error of 2.09%.Several experiments under various environment and weather were conducted as well,which indicated its robust performance with illumination changing.While in the size measurement tests,it showed a maximum range of 10 meters and average error in length and height were 3.24%and 2.52%,separately.In summary,this article was based on density map of point cloud and density-compensation algorithm to measure both the distance and the size of obstacles in the vicinity of camera.Unlike conventional image processing methods,it converted three dimensional point cloud data to two dimensional,using density of point clouds and applying grid map to significantly decrease the calculation amount.Besides,it functioned well under different weather conditions,indoor and outdoor environment,showing a robustness over traditional methods that separate obstacles from background in image processing.Whereas there were still some deficiencies to be improved,the current method and programming platform are still too time-consuming to fulfill the demand of real-time detection,it would be the future study direction in this field.