21 JUL 2026 (TUE) 11:05 - 11:35

Understanding Taxi Crashes in Hong Kong by Integrating Spatial Information, Visual Intelligence, and Behavior Dynamics

Mr. ZHUGE Bin    

( Supervisor: Prof Becky P Y Loo )

Abstract:

Road safety is a critical global public health issue, with traffic crashes causing substantial human, social, and economic losses. Within this broader road safety challenge, taxi safety deserves particular attention because taxis operate for long hours, carry passengers, and are continuously exposed to complex urban traffic environments. In Hong Kong, taxi-involved crashes have shown an upward trend in recent years, highlighting the need for a more systematic understanding of taxi crash risk. Although previous studies have examined general traffic crashes and selected risk factors, existing research often focuses on isolated factors and has not sufficiently integrated spatial information, visual intelligence, and behavior dynamics within a unified analytical framework. 

This study proposes two objectives: First, to understand taxi crash risk in Hong Kong by integrating spatial information, visual intelligence, and behavioral dynamics factors. Second, to investigate how the effects of crash risk factors on taxi crashes differ between daytime and nighttime conditions. To achieve these objectives, this study integrates multiple data sources, including annual taxi crash records, open spatial datasets, Google Street View images, and taxi dashcam videos. Spatial information is extracted through spatial data processing and spatial joints to identify crash hotspots and hot zones and characterize the surrounding urban environment. Visual intelligence is derived from Google Street View images and dashcam videos using computer vision methods. Behavior dynamics are extracted from dashcam videos and GPS data through video analytics, including object detection, vehicle and pedestrian counting, lane detection, overtaking identification, relative speed estimation, luminance-based measures, and visual entropy calculation. Based on these multi-source variables, machine learning models are developed to predict taxi crash risk, assess the relative importance of spatial, visual, and behavioral factors, and examine potential nonlinear relationships between risk factors and taxi crashes. 

The expected contributions are: First, it provides an integrated framework for analyzing how taxi crash risk is jointly shaped by where taxis operate, what drivers visually perceive, and how drivers behave dynamically. Second, by comparing daytime and nighttime crash mechanisms, this study reveals whether behavior dynamics contribute more strongly to nighttime taxi crash risk under reduced visibility, low illumination, glare, and changing exposure conditions. The findings are expected to support more targeted taxi safety interventions, including crash hotspots and hot areas management, street design optimization, nighttime lighting enhancement, and behavior-oriented safety management.