Assessing Effectiveness of Public Space Regeneration through Computer-vision-based Behavioural Analysis: Evidence from Daxianglu Street, Nanjing

Abstract

At the forefront of urban life, public spaces are constantly modified or even redeveloped to meet the needs of current lifestyle trends. How can to create better public space for everyone? It is a persistent and core question in planning and design practice for public space. This research focuses on streets - one of the most common public spaces – and uses a street regeneration project in central Nanjing as an example to explore the response to this question. Since public spaces accommodate a panoply of activities, their quality is unquestionably determined by their users, or simply, depends on how people use public spaces. Therefore, behaviour observation is a key approach to measuring the quality of public spaces (Kaparias et al., 2015, Motomura et al., 2022). Particularly, in the regeneration of public spaces, the comparative observation between the original environment and the renovated one can evaluate the effectiveness of its planning and design objectives.

Previous research often assesses public spaces by calculating behaviour-related indexes through observing recording and/or interviewing (Mehta and Bosson, 2021, Gehl and Svarre, 2013), and have been applied to assist regeneration design or to evaluate the regeneration’s achievements in vitality, safety and walkability (Kang, 2019, Kaparias et al., 2015). However, most of these studies based on public space activities are either time-consuming or limited in location inaccuracy. In recent years, the vigorous development of computer vision (CV) algorithms in behaviour recognition and trajectory tracking has made it possible to detect and analyse more detailed behaviours in public spaces more efficiently and accurately.

This research develops a new tool based on CV technology to automatically detect, trace and classify activities in open public spaces, at the precision level of decimetre. This precise large-scale data offers an opportunity to explore the relationship between human behaviours and spatial elements. This approach is applied to and tested in a regeneration project of Daxianglu Street, a deteriorated residential street in central Nanjing. Its relatively short duration of project planning, design and construction (less than six months) offers this research the chance to discover the relationship between spatial alteration and behavioural changes in analogous socio-economic settings.   

The natural experiment includes four steps in the case of the Daxianglu Street regeneration. First, unmanned aerial vehicles are used to capture videos on key street segments at the request of planners and designers. The videos are processed through CV technology (YOLOv5 and StrongSort++) to detect pedestrians, cyclists and cars and trace their trajectories. Second, the Random Forest machine learning algorithm is applied to classify pedestrians’ activities into eight categories depending on their status (standing, lingering, walking) and belonging group size (solo, small group, large group). Third, it provides precise and dynamic mapping and analysis of complex street activities based on the regeneration plans. Lastly, this research attempts to explore the relationship between spatial alterations and behavioural changes.

This research expects to offer a new approach to efficiently and objectively assess people’s activities in urban streets. This can help to discover the relationship between spatial elements and human behaviour in public spaces. The comparative data on various real regeneration projects have the potential to contribute to effective planning and design processes in future projects.