A Computer Vision System for Detecting and Analysing Critical Events in Cities

Whether for commuting or leisure, cycling is a growing transport mode in many cities worldwide. However, it is still perceived as a dangerous activity. Although serious incidents related to cycling leading to major injuries are rare, the fear of getting hit or falling hinders the expansion of cycling as a major transport mode. Indeed, it has been shown that focusing on serious injuries only touches the tip of the iceberg. Near miss data can provide much more information about potential problems and how to avoid risky situations that may lead to serious incidents. Unfortunately, there is a gap in the knowledge in identifying and analysing near misses. This hinders drawing statistically significant conclusions to provide measures for the built-environment that ensure a safer environment for people on bikes. In this research, we develop a method to detect and analyse near misses and their risk factors using artificial intelligence. This is accomplished by analysing video streams linked to near miss incidents within a novel framework relying on deep learning and computer vision. This framework automatically detects near misses and extracts their risk factors from video streams before analysing their statistical significance. It also provides practical solutions implemented in a camera with embedded AI (URBAN-i Box) and a cloud-based service (URBAN-i Cloud) to tackle the stated issue in the real-world settings for use by researchers, policy-makers, or citizens. The research aims to provide human-centred evidence that may enable policy-makers and planners to provide a safer built environment for cycling in London, or elsewhere. More broadly, this research aims to contribute to the scientific literature with the theoretical and empirical foundations of a computer vision system that can be utilised for detecting and analysing other critical events in a complex environment. Such a system can be applied to a wide range of events, such as traffic incidents, crime or overcrowding.