"Impact of Operational Ladar Occlusions on Point Cloud Instance Segment" by Andrew D. Gibson

Date of Award

6-2024

Document Type

Thesis

Degree Name

Master of Science

Department

Department of Operational Sciences

First Advisor

Bruce A. Cox, PhD

Abstract

Data exploitation techniques are the enabler for technological advancements in military ISR applications of ladar ISR. By identifying instances of military objects in observed scenes, point cloud deep learning models can unlock new standards of real-time information delivery to warfighters. Although current deep learning training datasets do not include real-world collection occlusions consistent with military applications, this research characterizes SPT model performance by adding occlusions to the DALESObjects dataset via artificial flyby simulations.We find that a baseline model trained on unoccluded data suffers performance degradation on both semantic and instance segmentation tasks when evaluated on occluded data, but that the effects can be minimized by collecting maximum angular coverage of the scene, especially at balanced observation elevations that stay away from extremely shallow or steep angles of collection. Beyond the most occluded edge cases, the model’s ability to identify vehicle instances within a scene was largely disconnected from the percentage of vehicle points observed. This suggests that identifying objects of interest within a scene can be done with a relatively sparse representation of that scene. Specialized models trained on occluded data exhibit increased performance on representative data compared to the baseline model. Matching training data to testing data maximizes performance, with degradations as testing data diverges in elevation or decreases in angular coverage. Increasing angular coverage can result in increased model performance but not above the performance of a network specially trained on that level of angular coverage.

AFIT Designator

AFIT-ENS-MS-24-J-030

Comments

A 12-month embargo was observed for posting this work on AFIT Scholar.

Distribution Statement A, Approved for Public Release. PA case number on file.

Share

COinS