Date of Award
5-2024
Document Type
Thesis
Degree Name
Master of Science in Engineering (MSE)
Department
Computer Engineering
Committee Chair/Advisor
Adam Hoover
Committee Member
Yongkai Wu
Committee Member
Xiaolong Ma
Abstract
This thesis investigates the problem of identifying traversable terrain in outdoor conditions. We are motivated by research in recent years toward identifying drivable space for the purpose of developing autonomous vehicles. Our motivating application is similar but also different. We envision a “Hiker Helper” that assists humans with dismounted navigation in forested terrain. A common challenge in this type of environment is identifying a viable path for moving through terrain that is congested with trees, bushes, other flora, and natural obstacles that would make navigation difficult. We envision training an artificial intelligence (AI) model to automatically analyze images of this type of terrain to identify potentially viable paths. The tool could highlight these areas in a camera view to help the user with navigation. Towards this goal, this thesis attempts to define traversable space based on single person dismounted navigation (e.g. hiking). This problem is more challenging than identifying drivable space for on-road navigation because the terrain appearance is much more variable. The terrain is not engineered to provide visible cues for navigation.
The work done in this thesis is intended to serve as a pilot project to aid in the study of this subject. One of the largest obstacles for this topic and for developing AI models is the lack of data and the quality of the data. The main results coming out of this project is a data annotation program that has been designed for users to quickly label data, definitions for labeling data, and a trained dataset. Use of optical flow was investigated thoroughly to determine its usefulness in helping semi-automate the process of labeling video frames by using a labeler’s input and interpolating their annotations to other frames. The annotation program is designed to allow users to label data with greater speed than other methods that exist, so that training data for this highly variable problem can be created. The dataset contains about 2.4 hours worth of videos in 30 different locations with 260,169 labeled frames. The ground truth for the videos consists of line labels that detail a traversable path that a hiker can take. These line labels were then used to create polygons that highlights a trail ii for pedestrians to walk through. Two independent annotators timed their labeling process for 80% of the data. The inter-rater reliability was measured by creating polygons that represented paths from the line labels created in the annotation program and then calculating the intersection-over-union, or IoU. The IoU was determined by dividing the overlapping area by the total area of the two paths. The inter-rater reliability between the two annotators was recorded to be 60.6% with an average label time being 0.40 and 0.31 seconds per frame for our two annotators.
Challenges for this project included determining exactly what kind of terrain was considered traversable and which direction to choose given multiple options. Issues such as lighting and motion blur in the video taking process disrupted the semi-automated labeling process by slowing or making the labeling process more complex. Future solutions that build on this thesis should strive to improve our definition for traversability which may make the labeling process more complex but may also increase inter-rater reliability. Future work could also try to speed up the labeling process by improving our video taking method for easier labeling or exploring more methods to alleviate the issues caused when optical flow is disrupted. The most important future work that could be done is to use the data set to train an AI that will automatically identify trails in an outdoor environment.
Recommended Citation
Nguyen, James, "Defining and Labeling Traversable Space in a Forested Environment" (2024). All Theses. 4284.
https://open.clemson.edu/all_theses/4284