Monocular Depth Estimation for 3D Scene Completion in Autonomous Racing

Wei Xun Lai

EECS Department
University of California, Berkeley
Technical Report No. UCB/EECS-2025-109
May 16, 2025

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-109.pdf

This thesis presents an approach to 3D scene completion for high-speed autonomous racing through monocular depth estimation. The research addresses a 22-degree LiDAR blindspot in the Indy Autonomous Challenge (IAC) racing platform, where vehicles operate at speeds exceeding 180 mph (290 km/h). At such speeds, comprehensive 360 degree environmental perception is paramount for safety and competitive performance, yet, due to engineering challenges, sensor configurations present significant gaps in coverage. To bridge this perceptual gap, we develop a transformer-based monocular depth estimation pipeline capable of cross-view generalization from front-facing to rear-facing perspectives without direct rear-view training data. Our approach extends reliable depth estimation from the conventional 40-60 meter range to 100 meters—providing an 80\% increase in reaction time for tactical decision-making at racing speeds. Extensive evaluation on the Las Vegas Road Course demonstrates robust cross-track generalization with a 28\% reduction in error variance compared to baseline approaches. The implementation is capable of achieving 42Hz inference throughput on the compute-constrained racing platform through TensorRT optimization and FP16 quantization. This work establishes monocular depth estimation as a viable complement to existing perception systems in autonomous racing, addressing sensor blindspots while maintaining the computational efficiency essential for real-time operation in competitive racing environments.

Advisor: S. Shankar Sastry

"; ?>

BibTeX citation:

@mastersthesis{Lai:EECS-2025-109,
    Author = {Lai, Wei Xun},
    Title = {Monocular Depth Estimation for 3D Scene Completion in Autonomous Racing},
    School = {EECS Department, University of California, Berkeley},
    Year = {2025},
    Month = {May},
    URL = {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-109.html},
    Number = {UCB/EECS-2025-109},
    Abstract = {This thesis presents an approach to 3D scene completion for high-speed autonomous racing through monocular depth estimation. The research addresses a 22-degree LiDAR blindspot in the Indy Autonomous Challenge (IAC) racing platform, where vehicles operate at speeds exceeding 180 mph (290 km/h). At such speeds, comprehensive 360 degree environmental perception is paramount for safety and competitive performance, yet, due to engineering challenges, sensor configurations present significant gaps in coverage. To bridge this perceptual gap, we develop a transformer-based monocular depth estimation pipeline capable of cross-view generalization from front-facing to rear-facing perspectives without direct rear-view training data. Our approach extends reliable depth estimation from the conventional 40-60 meter range to 100 meters—providing an 80\% increase in reaction time for tactical decision-making at racing speeds. Extensive evaluation on the Las Vegas Road Course demonstrates robust cross-track generalization with a 28\% reduction in error variance compared to baseline approaches. The implementation is capable of achieving 42Hz inference throughput on the compute-constrained racing platform through TensorRT optimization and FP16 quantization. This work establishes monocular depth estimation as a viable complement to existing perception systems in autonomous racing, addressing sensor blindspots while maintaining the computational efficiency essential for real-time operation in competitive racing environments.}
}

EndNote citation:

%0 Thesis
%A Lai, Wei Xun
%T Monocular Depth Estimation for 3D Scene Completion in Autonomous Racing
%I EECS Department, University of California, Berkeley
%D 2025
%8 May 16
%@ UCB/EECS-2025-109
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-109.html
%F Lai:EECS-2025-109