Inspiration

The inspiration for Invisible City Engine came from a simple but critical observation: road infrastructure is almost always improved after something goes wrong. Accidents, injuries, and damage are treated as the primary signals for action, while the countless unsafe situations that occur every day without resulting in a crash remain invisible. In other safety-critical domains such as aviation and industrial systems, near-misses are considered some of the most valuable sources of information. In road infrastructure, however, this perspective is still largely missing.

We were motivated by the question: what if cities could detect danger before accidents happen, instead of learning only after failure? This led us to explore how human behavior - hesitation, evasive maneuvers, abrupt braking, and avoidance - could be interpreted as early indicators of infrastructure-induced risk. Rather than adding more sensors or surveillance, we wanted to leverage signals that already exist but are currently ignored.

What we learned

Throughout the project, we learned that many of the most critical safety signals are subtle and distributed. Individually, a single near-miss or hesitation event may appear insignificant, but when aggregated spatially and temporally, clear patterns of latent risk emerge. This reinforced the idea that road safety cannot be fully understood through isolated events, but rather through behavioral trends that reflect how users interact with infrastructure under real conditions.

We also learned the importance of interpretability and operational relevance. Producing complex analytics is not enough, insights must be presented in a way that allows infrastructure operators to understand why a location is risky and how it can be improved. Balancing analytical depth with clarity became a central design principle of the project.

How we built the project

The project was built as a modular analytical pipeline that transforms aggregated mobility data into actionable infrastructure intelligence. We started by defining a set of pre-failure indicators, including near-miss events, hesitation patterns, evasive maneuvers, flow instability, and systematic avoidance. These indicators were extracted from simulated and voluntary mobility data using feature engineering techniques applied to speed, acceleration, heading, and temporal movement patterns.

On top of this feature layer, we designed composite risk indices that quantify latent infrastructure risk at the level of road segments and intersections. The results are visualized through a geospatial dashboard that highlights emerging hotspots, explains the contributing factors behind each risk score, and allows comparison across locations. A simple intervention simulation module was added to demonstrate how preventive measures could influence risk indicators over time.

Challenges we faced

One of the main challenges is working with noisy and heterogeneous mobility data while maintaining privacy-by-design. We address this by focusing on aggregated features rather than raw trajectories and by prioritizing robustness and explainability over overly complex models. Another challenge is defining meaningful risk metrics without relying on historical accident labels, which required careful reasoning about what constitutes a reliable pre-failure signal.

What's next

Invisible City Engine is designed as a foundation rather than a finished product. Future work will focus on validating the approach with real-world pilot data, refining the models using machine learning techniques, and integrating the system with existing traffic management platforms. The ultimate goal is to help road infrastructure evolve from a reactive system into one that can anticipate risk and act before failure occurs.