Robust Collaborative Perception for Smart Cities and Low-Altitude Economy

Abstract

Collaborative perception enables autonomous vehicles and mobile robots to share multi-view sensing data for more complete environmental understanding. However, practical deployment faces critical challenges: limited bandwidth, high transmission costs, time-varying wireless channels, and constrained edge resources. Existing methods rely on task-agnostic compression that wastes bandwidth on perception-irrelevant details, proving inefficient and insufficiently robust for safety-critical scenarios.

This thesis develops a task-oriented communication framework for collaborative perception at the wireless edge. Rather than optimizing bit-level reconstruction fidelity, we answer three fundamental resource allocation questions: who should collaborate, what data should be transmitted, and when should transmission occur. The proposed frameworks maintain perception performance under strict resource constraints, validated through extensive hardware testbed deployments.

First, we present the Priority-Aware Collaborative Perception (PACP) framework that addresses the collaborator selection problem in connected autonomous driving. PACP introduces a BEV-match metric to quantify viewpoint correlation and identify complementary observations. It jointly optimizes priority-aware bandwidth allocation, submodular collaborator selection, and adaptive compression tailored to channel states. Experiments demonstrate up to 41.5% improvement in perception accuracy while reducing communication costs.

Second, we develop the Prioritized Information Bottleneck (PIB) framework that determines what each camera should transmit. PIB embeds priority weights—derived from sensing coverage and channel quality—directly into the information bottleneck objective, enabling important cameras to preserve higher-fidelity features while aggressively compressing redundant views. A distributed online learning scheme enables adaptive edge server selection. Hardware evaluations demonstrate an 82% reduction in communication costs and 17.8% accuracy gains over standard codecs.

Third, we propose the Robust Adaptive Collaborative Perception (R-ACP) framework, which addresses the when-to-transmit dimension by unifying data freshness, calibration drift, and channel reliability. R-ACP introduces the Age of Perceived Targets (AoPT) metric that captures both temporal staleness and sensing quality to prioritize transmissions. A channel-aware Re-ID-based self-calibration method mitigates mobility-induced misalignment, achieving 89.4% improvement in calibration accuracy. Priority-aware fusion networks filter corrupted features to maintain robustness under packet loss. Results show 25.5% accuracy improvement and 51.4% reduction in communication costs under harsh channel conditions.

Finally, we extend task-oriented communication to GPS-denied aerial navigation within the Low-Altitude Economy, where extreme bandwidth constraints (sub-100 kbps) and limited onboard computation pose unique challenges. The proposed Orthogonally-constrained Variational Information Bottleneck (O-VIB) addresses the what-to-transmit question for multi-camera UAVs through automatic relevance determination and orthogonality constraints that eliminate cross-view redundancy. A split computing architecture balances lightweight encoding on UAVs with heavyweight localization at edge servers. UAV experiments achieve sub-10-meter localization at bandwidths below 10 KB/s, yielding 42.1% error reduction over vanilla VIB, 62.6% over WebP, and 95% lower latency.

Overall, this thesis presents a unified task-oriented communication framework that systematically addresses who collaborates, what is transmitted, and when transmission occurs across diverse collaborative perception scenarios. Hardware validation spanning connected vehicles, multi-camera edge systems, mobile robots, and UAVs demonstrates that the proposed approaches enable robust, real-time collaborative perception for Smart City infrastructure and the Low-Altitude Economy while satisfying stringent resource constraints.

Date of Award	27 Apr 2026
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Yuguang FANG (Supervisor)