17 JUN 2026 (WED) 16:05 - 16:35

The characteristics of pollution in Hong Kong's complex urban environment based on multimodal mobile monitoring and computer vision

Mr. ZHANG Chenming    

( Supervisor: Prof Nicky Y F Lam)

Abstract:

Traffic-related air pollution has become a critical global concern, posing a severe health risk to residents in urban areas. In cities such as Hong Kong, deep street canyons trap emissions at pedestrian level, while the growing fleets of electric vehicles (EVs) and hybrid electric vehicles (HEVs) further contribute to hyperlocal heterogeneity in pollution patterns. Previous monitoring efforts primarily focused on cameras at designated monitoring stations and key traffic routes, neglecting real-time pollution data from near roads. Meanwhile, conventional fixed-site monitoring stations are sparsely distributed and fail to resolve these fine-scale variations, leaving a critical research gap in attributing real-time local pollution to specific vehicle activities, such as fuel type and operating state. A unified methodology for identifying pollution contributors at the vehicle activity level and monitoring dynamic pollution patterns has yet to be established. 

Based on local statistics and real-time vehicle operation data, a vehicle-based mobile monitoring platform that fuses synchronised visible light (VL) and thermal infrared light (TIRL) cameras, real-time pollutant sensors, and GPS trajectories is developed to establish a direct link between real-time, vehicle-level activities and local air pollution in urban micro-environments. The platform enables per-vehicle classification by fuel type—internal combustion engine (ICE), EV, or HEV—and by operating state (idling or moving), even under low visibility conditions where VL alone fails. A deep learning model jointly processes the dual video streams frame-by-frame to detect vehicles, assign these attributes, and quantify each detected vehicle's contribution to synchronously measured pollutant concentrations. Field campaigns are conducted across representative Hong Kong micro-environments, including street canyons, tunnels, open roads and highways. The campaigns are repeated along the predetermined route and scheduled across three typical daily time windows: morning peak hour, noon, and afternoon-to-evening peak. Each campaign generates a novel multimodal dataset of aligned VL-TIRL vehicle images with vehicle fuel-type labels. By pairing per-frame vehicle attributes to GPS-synchronised pollutant measurements, a predictive model is constructed that infers local air quality characteristics directly from visual traffic features. The preliminary detection results indicate that TIR-based features enable more accurate differentiation of pollution contributors, driven by the enhanced thermal signals during engine boosting. Through vehicle-based mobile monitoring, this research establishes a direct, quantitative pathway from real-world traffic dynamics to local pollution characteristics. By linking visual traffic features to synchronised pollutant measurements, the research is expected to establish a reference standard for attributing pollution to transient vehicle activities.