3D modeling for agricultural field inspection robot

Abstract

Through automated agricultural inspection, farmers can potentially achieve better productivity and accurately predict yields and crop quality. The cheapest device for collecting rich information about a crop is the video camera. This thesis focuses on the design, implementation, and evaluation of algorithms for extracting information about a pineapple crop from a monocular webcam fixed to a mobile inspection platform. The process begins with fruit detection and tracking in video sequence. I propose a 3D modeling algorithm that works with the tracking history to generate a 3D point cloud for each fruit using structure-from motion techniques in computer vision. The 3D modeling algorithm consists of two main procedures: 3D point cloud estimation and 3D fruit shape reconstruction. Once the 3D point cloud for a fruit is obtained, the fruit shape reconstruction method estimates the shape and size of the fruit using a robust ellipsoid estimation technique. A series of experiments shows that the method produces a reasonably accurate 3D field map despite only partial and occluded views of the fruit. The prototype is thus a promising step towards the realization of automated mobile in-field agricultural inspection systems.