Proactive Road Safety Intervention in Australia: Predicting Risky Driving Hotspots from Connected Vehicle Data

Transport agencies in Australia have long depended on crash reports to spot dangerous roads, a method that only reveals problems after harm has already happened. Researchers from the University of New South Wales turned to the growing stream of data from connected vehicles to see whether risky maneuvers can be identified before they lead to collisions. By analysing hard braking, harsh cornering and sudden acceleration events recorded across Greater Sydney, they built a system that flags Local Government Areas where dangerous driving clusters.

What You Need to Know

The study used telemetry from thousands of equipped vehicles covering a six‑month period in 2023. Each trip supplied timestamped g‑force readings, which the authors converted into three binary indicators: hard braking (> 0.6 g), harsh cornering (> 0.47 g) and harsh acceleration (> 0.5 g). When any indicator fired, the event was logged as a “near‑miss risky driving” incident.

These incidents were aggregated to the Local Government Area (LGA) level and visualised as spatio‑temporal heatmaps, revealing pockets where risky behaviour repeatedly occurred during peak travel times. Eight predictive models—including logistic regression, random forest, gradient boosting and a simple temporal naïve baseline—were trained on 70 % of the data and tested on the remaining 30 %. Performance was measured with area under the ROC curve (AUC) and precision‑recall F1 scores.

Why It Matters

Identifying high‑risk zones before crashes occur allows traffic engineers to intervene with targeted measures such as adjusted signal timing, enhanced signage, or temporary speed reductions. Because the approach relies on already‑collected vehicle data, it can be scaled to other cities without waiting for expensive infrastructure upgrades or relying solely on police‑reported crashes.

Moreover, the method provides a continuous safety metric that can be monitored in near‑real time, offering feedback for evaluating the effectiveness of any intervention. By focusing on behavioural proxies rather than outcomes, the framework shifts road safety from a reactive to a preventive stance.

Key Details

• Data source: connected vehicle telematics from Greater Sydney, ~1.2 million trips, six‑month window.
• Risky driving thresholds: hard braking > 0.6 g, harsh cornering > 0.47 g, harsh acceleration > 0.5 g.
• Spatial unit: Local Government Area (LGA), 38 LGAs covered in the study.
• Models evaluated: logistic regression, decision tree, random forest, XGBoost, LightGBM, CatBoost, feed‑forward neural network, temporal naïve baseline.
• Best performer: LightGBM achieved AUC = 0.84 and F1 = 0.78 on the hold‑out set.
• Temporal resolution: risk scores computed hourly, enabling detection of rush‑hour spikes.

What’s Next

The authors plan to extend the workflow to incorporate weather and road‑condition data, test the model in other Australian states, and pilot a live dashboard for transport authorities that updates risk hotspots every 15 minutes. They also intend to explore whether combining near‑miss signals with low‑severity crash reports improves prediction of serious incidents.

📌 Source: Arxiv Ml