Enhancing Design, Management, and Operation of Social Infrastructure Through Deep Learning Methods

Abstract

Social infrastructure (e.g., public parks, squares, markets, etc.) plays a pivotal role in bolstering community quality of life, promoting economic prosperity, public health, and community resilience. Building on the work of Jane Jacobs, research has shown that walkable, mixed-use neighborhoods with a higher concentration of social gathering places and public space encourage the development of social capital and place attachment through an increase in social interaction. In short, the built environment and social organization of people are intimately connected. While autonomous sensing systems are ubiquitous in the modern world for monitoring and managing other forms of critical infrastructure, none yet exist for social infrastructure. Furthermore, despite its profound impact on the social health and resilience of the communities it supports, social infrastructure is often underfunded and underutilized. While data surrounding social infrastructure use and how communities interact with and derive benefits from it would certainly improve design, management, and operation of social infrastructure and address issues surrounding funding and utilization such data is hard to come by. Autonomous solutions are too simple, yielding only traffic estimates of people and cyclists. In depth datasets are too expensive and tedious to collect relying on pen and paper approaches to manually observe and annotate how social infrastructure is being used. This research aims to push a complete paradigm shift in how we design, manage, and operate social infrastructure. This document presents a first of its kind fully autonomous sensing system to detect, track, and map persons as they engage with social infrastructure all the while classifying their social behaviors and quantifying the sociability of the space. With this system, stakeholders (social infrastructure managers, and operators) can rigorously assess performance of their space and comprehensively analyze the impact of investments, programming, and design choices on the social health and utilization of their spaces. The sensing framework presented utilizes image streams from surveillance cameras and includes a multi-object detection component reinforced with a mapping module which projects pixel coordinates of detected individuals to a world coordinate system. Additionally, a tracking module, using data from the object detector and mapping module, traces individual trajectories and tracks time on site and interaction with park infrastructure. A sociability module then analyzes each trajectory and annotates the social interactions and behaviors observed for each tracked person, classifying their behaviors according to a novel schema. The social classification schema builds upon existing classification systems in urban sociology with modifications coming from insights gleamed from a series of in-depth interviews with critical stakeholders on the concept of sociability and how it relates to public spaces. From these interviews desired measurable outcomes and social behaviors were identified, leading to the development of activity and social indices’’, calculable from outputs of the sociability module, to capture and quantify the performance of a space. The sensing system was designed, developed, and implemented on a series of community parks along the Detroit Riverfront. These parks were chosen to serve as the primary research sites due to the diverse park social programming run in the spaces as well as the rich infrastructure (e.g., food carts, carousels, playgrounds, etc.) featured throughout the spaces. The system generated detailed reports on space use and enhanced decision-making processes of critical stakeholders showcasing the impact it will have on community health and how shared spaces are designed and managed.