Theses and Dissertations

Issuing Body

Mississippi State University

Advisor

Bethel, Cindy L.

Committee Member

Carruth, Daniel W.

Committee Member

Goodin, Chris

Committee Member

Chen, Jingdao

Date of Degree

12-13-2024

Original embargo terms

Visible MSU only 1 year

Document Type

Dissertation - Campus Access Only

Major

Computer Science

Degree Name

Doctor of Philosophy (Ph.D.)

College

James Worth Bagley College of Engineering

Department

Department of Computer Science and Engineering

Abstract

Vegetation override is an important aspect of off-road ground vehicle mobility. An autonomous ground vehicle’s (AGV) perception system must distinguish between vegetation that can be easily driven through from vegetation that cannot. Predicting the resistance of vegetation could allow path- planning systems to make this distinction. However, despite its importance, direct measurement of vegetation resistance is rare, as most studies use indirect proprioceptive data, such as inertial measurements, as proxies for override force. Notably, there is a lack of empirical data on the override resistance of small stems (<2.5 cm) and clusters of vegetation on medium-sized (approx. 1000kg) vehicles. To address this gap, a comprehensive dataset of override measurements was collected for clumps of small vegetation relevant to intermediate-sized AGVs navigating off-road terrain. This dataset includes over 70 recordings using the Robot Operating System (ROS) during controlled driving experiments through small trees, grasses, and bushes. The collected data includes light detection and ranging (LiDAR) scans, imagery, force measurements from integrated load cells, and simultaneous localization and mapping (SLAM) information. A key contribution of this research is the development and calibration of a custom push bar system equipped with load cells to directly measure override forces. These measurements are compared to empirical models previously developed by the U.S. Army Corps of Engineers for larger single-stem vegetation. A preprocessing pipeline was developed to automatically extract and label LiDAR and camera data according to these force measurements. This self-labeled dataset was then used to train machine learning models that predict override resistance of vegetation from LiDAR and camera scans alone. This research characterizes the relationship between override forces and the observable features of vegetation as measured by LiDAR and camera sensors. Deep learning models were developed and trained to predict override forces based on different input modalities and features derived from point clouds and images. The performance of these models was compared across various input features to investigate how deep learning can create a generalizable and accurate force prediction system.

Download

Share

COinS