Back to search

University of Calgary

Model-Resource Co-Optimization for Collaborative Inference in VEC-Enabled Autonomous Driving

Abstract

dc:description.abstract

This thesis investigates the joint optimization of deep neural network (DNN) partitioning and radio/computing resource allocation for collaborative object detection in vehicular edge computing (VEC) networks that support autonomous vehicles (AVs). While state-of-the-art detectors such as YOLOv5 deliver high accuracy, their computational cost can exceed on-board budgets under strict latency targets. Collaborative inference, where early layers execute on the AV and later layers on a roadside unit (RSU), reduces on-board computa-tion and uplink payloads by transmitting intermediate features instead of raw images. We develop a two-timescale framework that (i) selects zone-specific model partition points at a slow timescale, and allocates, at each slot, RSU computing and bandwidth across zones to maximize the number of AVs whose end-to-end (E2E) inference latency meets a target. We formalize the problem as a mixed-integer nonlinear program (MINLP) that captures the coupling among computing times, intermediate feature size, and uplink rate via Shannon capacity. By eliminating the bandwidth variables at feasibility and smoothing the deadline indicator, we obtain per-slot convex programs over computing shares and embed them in a genetic algorithm (GA) that searches partition decisions. Experiments combine SUMO-generated traffic traces, a YOLOv5 detector fine-tuned on a consolidated BDD100K subset, and realistic wireless channels. Compared with greedy and partial-optimization baselines, the proposed GA+convex framework supports more AVs under a 100 ms latency target and approaches the performance of brute-force optimal search with substantially lower computa-tional cost. Ablations illustrate how partition points adapt to traffic density and link quality and show gains from finer-grained partition sets and feature-size reductions. We discuss de-ployment considerations (model versioning, quantization, fairness) and present guidelines for robust operation. Overall, the results demonstrate that coordinated model partitioning and resource allocation is an effective lever for meeting stringent perception latency service-level objectives in VEC-enabled autonomous driving.

Degree

thesis:*
Name thesis:degree_name
Master of Science (MSc)
Discipline thesis:degree_discipline
Engineering – Electrical & Computer
Year dc:date.issued
2025

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Qiu, Ye
Advisor dc:contributor.advisor
  • Wu, Huaqing
Committee members dc:contributor.committeemember
  • Drew, Steve
  • Hai, Jiang

Rights

dc:rights
Statement dc:rights
  • Unless otherwise indicated, this material is protected by copyright and has been made available with authorization from the copyright owner. You may use this material in any way that is permitted by the Copyright Act or through licensing that has been assigned to the document. For uses that are not allowable under copyright legislation or licensing, you are required to seek permission.
Language dc:language.iso
en

Identifiers

dc:identifier.*
Dc Identifier Citation
Qiu, Y. (2025). Model-resource co-optimization for collaborative inference in VEC-enabled autonomousdDriving (Master's thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca.

Chain of custody

source
Harvested from
University of Calgary
Base URL
ucalgary.scholaris.ca/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
related terms
citation

Qiu, Ye. Model-Resource Co-Optimization for Collaborative Inference in VEC-Enabled Autonomous Driving. 2025. https://hdl.handle.net/1880/123710