Biological Systems Engineering, Department of


First Advisor

Santosh Pitla

Date of this Version


Document Type



Lindhorst, C (2019) Field Obstacle Identification for Autonomous Tractor Applications


A THESIS Presented to the Faculty of The Graduate College at the University of Nebraska In Partial Fulfillment of Requirements For the Degree of Master of Science, Major: Agricultural and Biological Systems Engineering, Under the Supervision of Professor Santosh Pitla. Lincoln, Nebraska: April 2019.

Copyright (c) 2019 Caleb Lindhorst


New technologies are being developed to meet the growing demand for agricultural products. Autonomous tractors are one of the many solutions to address this demand. Obstacle detection and avoidance is an important consideration for safe operation of any autonomous machine. Three field obstacles were chosen to be identified in this thesis work: tractors, round bales, and center pivots. Limited research work was found on the identification of center pivot detection.

Feasibility of using low cost LIDARs was considered for the detection of tractors, bales, and agricultural center pivots. Performance of LIDARs in different lighting conditions, different colors of obstacles, accuracy and angular resolution was evaluated. It was found that low cost LIDARs do not have a small enough angular resolution to detect pivots at a distance to avoid the obstacle. Formulas were derived to help find the distance between steps of the LIDAR.

Obstacle identification is also important so that proper corrective actions can be taken to avoid the obstacle. RGB cameras were used to aid in the detection of center pivots. SURF Feature Extraction and Matching, Viola-Jones algorithm and edge detection with a shape identification algorithm were tried but none of the algorithms could adapt to more than one orientation or class of object.

Obstacle identification using Convolutional Neural Networks (CNNs) for obstacle detection was pursued. Each obstacle was individually trained first and then all classes were combined to create one object detector. Faster Region based CNN (R-CNN) was used with GoogLeNet to give high mean Average Precision (mAP).

Advisor: Santosh Pitla