Analytics for Better Urban Cycling

Abstract

Cycling has become an increasingly popular mode of transportation, offering significant benefits for urban mobility, public health, economic growth, and sustainability. However, urban infrastructure in North America has traditionally been car-centric. Safety and comfort concerns remain significant barriers to broader cycling adoption. As cities strive toward a net-zero future, redesigning urban spaces to encourage a shift toward this more sustainable transportation mode is essential. In line with this vision, this thesis develops quantitative methods for the efficient assessment, planning, and utilization of urban bike infrastructure, with a particular focus on fostering cycling adoption in Toronto, Canada. First, we introduce a computer vision approach to assess cycling stress---the discomfort cyclists experience on urban road networks. This method leverages the widespread availability of street-view images to replace current data- and labour-intensive assessment practices. Next, we formulate an optimization model to determine the optimal locations for new bike lanes, aiming to maximize the city's low-stress cycling accessibility---a metric strongly correlated with cycling mode choice in Toronto. Given the size and complexity of Toronto's road network, this optimization model cannot be solved by any existing methods. In response, we develop a machine learning-augmented optimization approach that is computationally efficient and produces provably high-quality solutions. From 2020--2024, this approach achieves comparable performance to Toronto's implemented plan while reducing the required length of bike lanes by 25%, equivalent to a cost saving of over 18 million Canadian dollars. Alternatively, the same infrastructure investment could have yielded an additional 11.2% increase in cycling accessibility. Finally, we investigate cycling path recommendations in the context of last-mile delivery. Empirical evidence indicates that cycling couriers often deviate from delivery routes prescribed by platforms, leading to increased delivery times and inefficiencies in order batching and assignment, both of which depend heavily on accurate delivery route modelling. To address this, we propose a data-driven approach to prescribing delivery paths that are both high-quality and perceived as such by couriers. The former property ensures service quality, while the latter promotes adoption. In a case study, our pipeline significantly improves route adherence without compromising delivery times compared to current industry practices.