Growth Parameters Extraction For Crop Seedlings In Greenhouse With RGB-D-Based

In order to provide appropriate temperature,humidity,sunlight and nutrients for caltivation seedings in greenhouse,the growth parameters which are key aspect for predicate the growth states is needed for precisely control growth environment.Recently,for the purpose of precisely planting seedlings,based on captured parameters in realtime to dynamic adjusts the cultivation condition become research hotspot.However,there are problems unsolved such as expensive data capture equipment,limited of detection objects and growth parameter,and some of methods are invasive to seedling.In our research,we aim to capture the growth parameter in approach of non invasive with the Kienct sensor.The main contributions of this thesis are as follows:(1)An experiment platform to capture seedling data in greenhouse with KinectV2 which connected to a PC workstation and erected in a tripod was presented.Our plantform can capture the depth and RGB data synchronously at frame rate of 30 fps,and captured datas meet the needs of the subsequent data fusion to improve accuracy.Compared to those professional sensors,our plantform is inexpensive and achive acceptable result.(2)To extracting the research object and filter noise such trays,tripods,scene facilities etc,we propused a pre-processing approach which combine of useable-area filter,depth cut-off filter,and neighbor count filter.The above three filter for the purpose of removing unusable area,unusable depth and other tiny noise respectively.Our pre-processing result showed that the combined three filters remove the noise effective,and provide reliable raw data for the subsequent research.(3)To segement the seedlings from tray which contains soil,a surface feature based segemention method was propused.The proposed method construct the feature vector which consists of point coordinates,Lab color space and point fearture in thirty-nine dimention.To clustering the points with similar feartures,we employ K-means algorithm and improved the calculation of feature difference by Hailinger distance.Compared with the 2g-r-b index and the improved SANSAC,our method overcomes over-segment and boresome parameter revision,and the segmenation procedure can automatic execute after setting the parameter.(4)In order to segment every seedling from the seedlings,the VCCS+LCCP algorithm is applied.VCCS+LCCP is an efficient learning-and model-free approach for the segmentation of 3D point clouds into object parts,the first step is over-segement the object,and use Voxel Cloud Connectivity(VCCS)to build surface patch adjacency graph(a 3D analog of superpixels).Then,segment the supervoxel adjacency graph by classifying whether the connection between two supervoxel is convex(= valid)or concave(= invalid).VCCS+LCCP algorithm effectively solve the problem that two closely spaced or overlapping seedlings in Euclidean distance clustering are clustered into the same cluster.In order to achieve the extraction of the key growth parameters of seedlings,the key growth parameter,namely the three growth parameters of leaf color,leaf area,and plant height,were extracted from the clustered 3D points of single vegetable seedlings.Afte implemented the above menthioned approach,our experiment result show that in the segmentation of single seedling,the identification accuracy for bean sprouts and young cabbage seedlings with regular leaf surface can be up to 100%;for the irregular lettuce seedlings,the identification accuracy can be up to 95%;for coniferous tomato seedlings,the identification accuracy can be as high as 72%.In the measure of the height of bean sprouts,the mean measure error of bean sprouts was 2.30 mm,and the mean relative measure error was 7.69%.This result can provide an effective reference solution for extracting the growth parameters of seedlings.