Back to search

Cornell University

Enhancing 3D Perception with Unlabeled Repeated Historical Data for Autonomous Vehicles

Abstract

dc:description.abstract

The evolution of autonomous vehicles is advancing rapidly, promising a radical shift in our future mobility. The cornerstone of building a reliable autonomous vehicle hinges on developing a robust perception system capable of accurately perceiving the vehicle's surroundings, primarily through the detection of other road participants. The current predominant approach relies heavily on supervised learning, requiring a substantial quantity of costly labeled data, and employs expensive 3D sensors such as LiDARs. This approach, while comprehensive, has intrinsic limitations: disambiguating small or distant objects becomes challenging due to the sparse nature of LiDAR point clouds. Furthermore, the supervised perception system is sensitive to domain differences when deployed in diverse environments, and cost-effective camera-only alternatives fail to achieve the accuracy of their LiDAR-based counterparts. These issues necessitate a critical reevaluation of the conventional methodology for 3D perception in autonomous vehicles. In this thesis, we investigated a novel yet highly practical strategy that could significantly mitigate these concerns — leveraging past visit data from the same location. Previous approaches tend to treat each scene as a new, unknown entity, ignoring the potential wealth of information from earlier traversals of identical routes. In reality, our daily commutes often involve repetitive journeys: to workplaces, schools, stores, and social gatherings. Even when navigating an unfamiliar route, we are typically retracing the paths of preceding drivers. We demonstrated that this simple shift in data perspective can yield substantial benefits: 1) Hindsight: a significant, consistent improvement in the performance of LiDAR-based 3D object detection models, 2) Rote-DA: a pronounced reduction in performance degradation of 3D perception models due to domain differences, and 3) AsyncDepth: a consistent improvement in camera-based 3D object detection models when supplemented with past LiDAR data. These promising results signify that historical data, a rich trove of information, can significantly contribute to improving 3D perception for safer autonomous driving. We also discussed potential future research directions at the conclusion of this thesis.

Degree

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

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • You, Yurong
Chair dc:contributor.chair
  • Weinberger, Kilian
Committee members dc:contributor.committeemember
  • Hariharan, Bharath
  • Sun, Wen
  • Bindel, David

Subjects

dc:subject × 5

Rights

dc:rights
Statement dc:rights
  • Attribution 4.0 International
Language dc:language.iso
en

Identifiers

dc:identifier.*
Dc Identifier Bibid
16219328
Dc Identifier Other
ProQuest Submission ID: 13907
ProQuest Publication ID: 30576091

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

You, Yurong. Enhancing 3D Perception with Unlabeled Repeated Historical Data for Autonomous Vehicles. Doctor of Philosophy thesis, Cornell University, 2023. https://hdl.handle.net/1813/114814