Study On Plant 3D Reconstruction And Information Acquisition Under Different Environment

China is the top one agricultural producing country in the world,with the most yield of vegetables.That’s the reason,why we need to monitor the crops information to improvement growth environment and management.The Kinect V2 and high resolution RGB action camera were used in the study to acquired plant data.Aimed at plant weight accurated evaluation.Exploring the practical and effective working pipeline to reconstruct plant 3D model for plant monology parameters computing,in different growth environment.Based on the unique imaging environment,this study designed and optimized different imaging acquiring and reconstruction methods.A deep-learning based real-time counting methods for maize seedlings is developed in this study also.This study do relevant researches from the following four parts:(1)Plant growth parameters acquisition in lab with depth cameraBased on our research group previous study,this study proposed a plant reconstruction method.Top-view RGBD images were captured with Kinect V2 camera in lab to reconstruct plant3 D model.A leaf template based method was proposed to fixing the incomplete pointcloud of vegetables,which was caused by fieldview occlusion.After fixing the plant pointcloud,we can extracte accurated plant growth parameters.(2)Plant growth parameters acquisition in greenhouse with depth camera.Promoting the plant reconstruction and parameters acquirement method form lab to greenhouse.Optimizing the method according to the data acquiring environment.After acquiring segmented plant,we used the extracted parameters method in parts(1)to retrieval plant height and total area.After linear fitting analysis,plant height correlation analysis with manually measured plant height is 0.8234,correlation analysis between plant total area and plant weight reached 0.938.(3)Plant growth parameters acquisition in plant factory with depth and RGB cameraConsidering the planting environment in plant factory we combined RGB camera and Kinect V2 depth camera to collect plant data.As for the hole in plant point cloud with Structure From Motion reconstruction method,we mended holes manually and traind the MSN(Morphing and Sampling Network)pointcloud completion network,which could fix holes automatically.And after pointcloud completion,we calculate the whole planting panel total area to estimate weight for the single planting panel.The correlation between the total area and wight for single planting panel is 0.8271 for Regiment spinach and 0.8971 for Savoy spinach(4)Plant population acquisition in field with RGB cameraThis part proposed a corn seedlings real-time couning based on 2D inage.A action camera was mounted on a cart to capture corn seedlings video from top view.A counting Pipeline based on Yolo V3(You Only Look Once,Version 3)and Kalman filter was proposed to detecte and count corn seedlings.This counting method can reach 98% accuracy in stage V2 and V3.