Predicting and Preventing Roadside Tragedies

Client information / Background

In order to reduce road incidents, a government department focusing on traffic management and road safety, sought to innovate their approach. Traditionally reliant on analysing historical data for insights into past accidents, they recognised the need for a predictive model to forecast future hotspots and enable proactive prevention strategies. Such advanced analytics were not supported by their existing infrastructure. Aimed at moving from reactive to proactive traffic management, the department made use of  Spatialedge consulting and delivery services to leverage big data and AI, targeting a significant reduction in road accidents and fatalities.

Problem

They approached our team with a challenge: they possessed extensive data on historical accidents but lacked predictive insights into future accident hotspots. Their existing infrastructure included basic analytics dashboards for reviewing past incidents. That is a reasonable indicator, but there are just too many areas to cover with the limited road safety vehicles available. Therefore the problem each municipality faced was, “where should we spend time and effort with road safety initiatives to reduce fatalities”?

Solution

To solve this problem requires advanced analytics. Picking the spots most likely to have accidents and then targeting those areas with the limited resources.

Data Processing

The initial phase encountered significant challenges due to the lack of structured data pipelines and databases. The team was provided with raw data file drops, necessitating the construction of data processing workflows from the ground up. A significant portion of this phase involved geocoding loosely described accident locations into precise latitude and longitude coordinates. This was achieved by cross-referencing accident location descriptions with telemetry data from forensic pathology services vehicles, enabling accurate identification of accident locations.

Modelling Approach

The modelling began with kernel density estimation to identify accident likelihood based on historical data. This approach was soon replaced by spatial indexing in conjunction with machine learning classifiers to incorporate additional variables such as road network characteristics  and weather conditions. The model divided the map into segments, each with its own set of features, to predict accident occurrences. This evolved approach allowed for a more dynamic understanding of accident hotspots, considering various influencing factors.

Outputs

The project produced a predictive model capable of identifying future accident hotspots with considerable accuracy. These outputs were then made actionable through the development of a prescriptive model, which recommended optimal locations for traffic law enforcement operations. Additionally, a root cause analysis tool was developed to assist in planning and to help create educational campaigns, particularly focusing on pedestrian safety.

Implementation and Integration

The lack of integration with the client's existing IT infrastructure posed challenges, leading the team to operate within a siloed environment initially. Despite these hurdles, the outputs were integrated into a user-friendly dashboard, providing stakeholders with actionable insights for planning and operational decisions.

Result

The project yielded measurable improvements in road safety outcomes. An example metric shared was a 30% reduction in fatalities over a major public holiday weekend, compared to the previous year. However, challenges in directly attributing these improvements to the project's outputs were acknowledged, given the complexity of factors influencing accident rates. 

Future work was recommended to further refine the models, add additional features, and explore deeper integrations with client systems to operationalise the findings more effectively.