Research On Plant Point Cloud Information Fusion Method Based On Kinect And Laser Sensor

The acquisition and processing of 3D point cloud information are widely used in side-view monitoring activities of high row-cultivated plants such as applications of pesticides,irrigation,fertilization,crop training and harvesting.The information fusion of plant point cloud is a hot spot of agricultural intelligence and precision.The information of spatial,luminance,intensity and color can be complemented through the fusion of different angle sources of the single sensor or multi-sensor information fusion,to enhance the 3D point cloud reconstruction information.To solve the problem of low precision and slow speed of the traditional point cloud registration method.In his paper,a method of improved SIFT-ICP plant point cloud registration based on Kinect is presented,which show a better result of registration in speed and accuracy.Furthermore,to avoid the edge missing of the target image and the influence of black hole,the SICK laser sensor is used to obtain the 3D point cloud information of the plant.The point cloud interpolation method of Kriging plant based on super-body clustering is studied,which makes the plant level richer and the texture clearer.Finally,the information of both sensors was combined,which take advantages of the laser sensor,which is not disturbed by the sun,and Kinect which can acquire color and depth information simultaneously.In this theses,fusion information fusion based on SICK and Kinect combination detection method is proposed.The main contents of the theses are as follows:(1)Because of the low precision and slow speed of the traditional point cloud registration,a new method of Kinect plant point cloud registration based on improved SIFT-ICP is proposed,and the fast and accurate registration of plant point clouds in different perspectives is realized by using Kinect sensors.First,the point cloud method is preprocessed to remove the background and noise and extract the point cloud of the target plant,then we carry out the SIFT key point search method.The adaptive normal line estimation,the fast point feature histogram is calculated,secondly;SAC-IA initial registration was carried out to realize the initial change of the point cloud space position;finally,the ICP algorithm,accelerated by Nanoflann,is used for accurate registration.In the experiment,24 point clouds have been obtained from different angles.First,the point cloud has been pretreated to remove the background noise,and then only the plants point clouds were obtained from point cloud;lastly,the registration was performed.Compared with the traditional ICP,PCA-ICP and SIFT-ICP registration.The registration errors have been reduced from 4.23 cm,2.53 cm and 1.41 cm to 0.48 cm respectively.The registration accuracy has been increased by 88.7%,81% & 66% respectively.The registration time decreases from 56.2s,189.5s,138.6s to 26.6s,and the percentage of registration time can be represented as 52.7%,86% and 80.8% respectively.To verify the stability of the test,22 registration of point clouds with the interval of 30 degrees from the plant angle of view has been considered.The results show that the registration error always kept within 0.7cm,and the registration time was not more than 30 s.Considering the edge loss of Kinect caused by sunlight interference,an outdoor supplementary experiment has been carried out for further confirmation of the proposed algorithm.(2)Given the phenomenon that the point cloud reconstruction of SICK sensor is sparse during the moving process and the traditional point cloud interpolation is not good,a Kriging plant point cloud interpolation method based on the hyperbody clustering is proposed.The partition and interpolation method is firstly used to realize the partition and block interpolation of point cloud.Firstly,the global segmentation based on hyperbody clustering is carried out.Because the interior of the point cloud contains more objects,complex environment and the phenomenon of object stacking,global segmentation can realize the segmentation of different object point clouds,and the problem of point cloud interpolation confusion between objects are solved.Then,the internal segmentation of single objects based on LCCP is carried out for different objects,so that different concave and convex points can be segmented.The internal segmentation of single object can effectively comb out the internal structure of the object and simplify the complex structure,which can effectively reduce the difficulty of interpolation.Finally,Kriging interpolation is carried out for different point cloud regions,and precision interpolation of small area points is realized by using the precision of Kriging interpolation algorithm.After interpolating each cell domain,the global interpolation of point clouds is realized.The experimental result shows that the number of laser point clouds after interpolation is increased,the level is richer,and the texture is clearer.(3)In view of the problem that Kinect point cloud contain the edge missing and black hole when working outdoor and traditional point cloud information fusion need complex calibration before working,and keep the location of the sensor,a fusion method based on SICK and Kinect combined detection is proposed.The point cloud obtained by the SICK sensor is reconstructed,then segmented and interpolated,and the point cloud obtained by Kinect is preprocessed,the target plants in the two-point clouds are extracted and registered.In order to accelerate the registration process,the two points are down-sampled under the voxel grid.With the initial registration and accurate registration,the location of the point cloud after registration is closed,and the nearest neighbor search is used to find the corresponding point of SICK point cloud in the Kinect point cloud.In order to improve the accuracy of information fusion,a method of exceeding the limit compensation is designed for further fusion,which effectively corrects the inaccurate point cloud color and empty holes.According to it.The experimental results show that the similarity between the fused point cloud and the original point cloud image is more than 90%,and has a high accuracy,which solves the disadvantage of the limited outdoor of Kinect.