Abstract: Aiming at current difficulties in non-contact fast estimation of the weight of non-detached tomato fruits，a method for estimation of tomato fruit weight based on local point cloud and convolutional neural network was proposed.Taking ZheFen 702 tomato as experimental object，firstly，336 original point clouds of 50 single tomato fruits were collected by a depth camera，and then enhanced to 1 344 point clouds for constructing a data set.A variety of point cloud segmentation methods were compared and selected，three-dimensional continuous convolutional neural network was used for tomato single fruit segmentation.Based on segmented data，a total of 4 spatial feature information point cloud were extracted，which were axial size dx along x axis，axial size dy along y axis，a axial size dz along z axis and minimum circumscribed circle diameter d along z axis projection.They were fed to three-layer regression network for training，determining optimizer and weight estimation model when learning rate reached optimal condition.Finally，268 enhancements point cloud were selected to test constructed mathematical model，evaluate and analyze accuracy and stability of the model.Results showed that compared with actual weight of a single tomato，average deviation was 3.7~4.8 g，and average relative error was about 3.04%，which was better than traditional image processing methods.This research could provide technical reference for non-contact weight estimation of other agricultural and livestock products.