Back to search

Cornell University

Active Multiview Perception for Mobile Sensors

Abstract

dc:description.abstract

Imaging sensors, including sonar, electro-optical sensors, and LIDAR, are broadly used in mobile robotics for object recognition and 3D reconstruction tasks. Because only part of an object may be visible to a sensor at any given time, successful reconstruction and object classification often requires gathering and fusing images from multiple disparate viewpoints. Mobile robot operations are often energy and time-constrained. Therefore, imaging viewpoints must be planned judiciously to improve target sensing performance while avoiding time-consuming operations. This dissertation presents the development of several novel multiview target classification, reconstruction, and sensor path planning algorithms for imaging sensors. Specifically, a recursive Bayesian estimation method is used for multiview underwater target classification. The expected confidence level (ECL) metric is developed to select the most informative future viewpoints for the sensor. Experiments show that informative path planning algorithms using ECL outperform existing path planning methods for underwater multiview target classification. To address the issue of limited data pertaining to underwater environments, this dissertation introduces a novel method to generate underwater object imagery that is acoustically compliant with that generated by side-scan sonar using the Unreal Engine. This method provides visual approximations for acoustic effects such as back-scatter noise and acoustic shadow while allowing fast rendering with C++ actors in UE for maximizing the size of potential ATR training datasets. In addition, this dissertation presents a novel framework leveraging a generative model for energy-efficient, physics-constrained multiview target reconstruction and sensor path planning. Using an existing generative model to predict or complete an object geometry online future sensor views can be predicted, evaluated, and optimized subject to time and energy demands. Simulation results show that this approach outperforms state-of-the-art algorithms, such as a reinforcement learning approach, ScanRL. It is also shown that the estimated ground truth from the generative model enables other functionalities of the active reconstruction algorithms, such as anomaly or fault detection via visual inspection.

Degree

thesis:*
Name thesis:degree_name
Ph. D., Mechanical Engineering
Level thesis:degree_level
Doctor of Philosophy
Discipline thesis:degree_discipline
Mechanical Engineering
Grantor
Cornell University
Year dc:date.issued
2023

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Chang, Shi
Chair dc:contributor.chair
  • Ferrari, Silvia
Committee members dc:contributor.committeemember
  • MacMartin, Douglas
  • Weinberger, Kilian

Subjects

dc:subject × 6

Rights

Language dc:language.iso
en

Identifiers

dc:identifier.*
Dc Identifier Bibid
16176612
Dc Identifier Other
ProQuest Submission ID: 13486
ProQuest Publication ID: 30250583

Chain of custody

source
Harvested from
Cornell University
Base URL
ecommons.cornell.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Chang, Shi. Active Multiview Perception for Mobile Sensors. Doctor of Philosophy thesis, Cornell University, 2023. https://hdl.handle.net/1813/113999