A New Approach of Monocular Visual Odometry to Trajectory Estimation Within a Plantation

• Gabriel L. Araujo USP
• Jorge Id F. Filho USP
• Vitor A. H. Higuti EarthSense
• Marcelo Becker USP

Resumo

A projection that has worried many researchers is the lack of food due to population increase, generating a demand for more efficient methods of food production while also minimizing environmental footprint. To increase efficiency and reduce both costs and time in crop scouting, the application of autonomous robots for phenotyping research and monitoring of plantations is very attractive. In such context, this research proposes a new Visual Odometry (VO) approach to allow vision-based mapping of a area by a small robot within rows of an agricultural field. The proposed method estimates the robot’s pose from images treated by filters and color separation, binarization, edge detection, and Hough transform, where the latter returns the line, which is used to estimate the angle of a rotation matrix. The difference between computed angles from two frames is returned to the classical odometry visual algorithm for motion estimation. VO and the proposed method are able to estimate a trajectory in a straight line with small translation errors, at the final point of the path, in relation to the ground truth. The proposed method reaches smaller translation errors than VO and performs a path recuperation with fewer oscillations.

Palavras-chave: Visualization, Uncertainty, Motion estimation, Robot vision systems, Estimation, Transforms, Cameras