Computationally Efficient Autonomous Vehicle Perception: Intelligent Infrastructure, Algorithms, and Real World Applications

Date of Award

4-2025

Degree Name

Doctor of Philosophy

Department

Mechanical and Aerospace Engineering

First Advisor

Richard T. Meyer, Ph.D.

Second Advisor

Zachary D. Asher, Ph.D.

Third Advisor

Guan Yue Hong, Ph.D.

Fourth Advisor

Ali Riza Ekti, Ph.D.

Keywords

Autonomous vehicle, computational offloading, infrastructure information sources, LIDAR-camera fusion, perception

Abstract

Safe autonomous vehicle (AV) operation depends on having an accurate perception of the driving environment. Advanced computational algorithms are typically utilized to fuse the incoming data from various sensors for a robust perception and localization system, which historically demands a high computational load and associated large power usage. The high computational load and subsequent power usage leads to an increase in energy consumption, reducing vehicle range and raising operational costs. Three pivotal strategies are presented to reduce the computational workload and power usage in AV perception. First, the integration of Vehicle-to-Infrastructure (V2I) information exchange is investigated to enhance perception using contextual data from infrastructure information sources (IISs), such as chip-enabled raised pavement markers (CERPMs). These IISs improve perception accuracy while reducing onboard processing demands. For V2I integration, CERPMs were developed to enhance AV perception by providing lane line information. These CERPMs, when fused with camera data, were able to accurately estimate lane lateral offset while reducing the computational load.

Second, a computationally efficient deep-learning-based sensor fusion method for 3D object detection is developed. The presented approach, termed AEPF (Attention-Enabled Point Fusion), involved creating energy-efficient fusion methods based on deep learning that fuse camera images and LIDAR point clouds while maintaining state-of-the-art (SOTA) accuracy without imposing heavy computational demands. AEPF uses images and voxelized point cloud data as inputs and estimates the 3D bounding boxes of object locations as outputs. An attention mechanism was introduced to an existing feature fusion strategy to improve 3D detection accuracy and two variants are proposed. AEPF-Small, with a lightweight attention module and fewer parameters, offers 1.83 times faster inference than compared baseline. AEPF-Large, with a more complex attention module and increased parameters, provides higher mean average precision (+1.63) than the compared baseline.

Third, the potential of cloud computing for reduced computational load and associated energy use onboard the vehicle is investigated. An optimal scheduling scheme for cloudbased computational offloading is formulated to reduce the AV’s onboard compute load. The presented approach enabled energy-efficient AV operation while ensuring the timely return of processed data from the cloud. To examine cloud-based computational offloading, an optimal scheduling strategy for autonomous driving tasks via the cloud layer was formulated as a mixed-integer linear programming (MILP) problem and verified using synthetic task graphs. It is shown that cloud-based computational offloading enables energy-efficient AV operation while maintaining task deadlines to ensure the timely availability of processed information. Simulation results demonstrate that on average, 35.78% of the computational load was offloaded to the cloud, achieving 35.65% energy savings for the onboard system.

Together, these strategies take a step toward reducing energy usage for AV operations by investigating computational load reduction pathways. The broad impact is a more sustainable and energy-efficient approach to AV operation.

Access Setting

Dissertation-Abstract Only

Restricted to Campus until

5-1-2027

This document is currently not available here.

Share

COinS