Source-linked AI summary
Proactive Road Safety Intervention in Australia: Predicting Risky Driving Hotspots from Connected Vehicle Data
Adriana-Simona Mihăiţă, Clarence Cheung, Artur Grigorev, Tuo Mao, David Lillo-Trynes
TL;DR
Reactive crash-based monitoring leaves proactive identification of risky driving hotspots underexplored. Using connected-vehicle telemetry and LGA-level forecasting, the study finds ARIMA achieves the lowest MAE while remaining comparable to LSTM.
Problem
Proactive identification of high-risk locations and dangerous driving behaviour before incidents occur remains underexplored despite predominantly reactive safety interventions.
Source: §1.1 · p. 1Method
The study aggregates g-force-defined risky driving events from connected-vehicle telemetry at LGA level and benchmarks eight models across ensemble, deep-learning, and time-series families.
Source: §2 · p. 3, §3 · p. 8, §5 · p. 12Results
ARIMA achieved the lowest MAE (162.21), performed comparably to LSTM (163.92), and outperformed all ensemble learning methods.
Source: §4 · p. 11, §5 · p. 12Takeaways & Limitations
Consistently high-risk CBD, Parramatta, and Bankstown LGAs provide a spatial basis for targeted road safety interventions.
Source: §5 · p. 12Daily aggregated observations were insufficient for data-hungry deep-learning architectures, and evaluation covered only the final 35 days of each LGA series.
Source: §5 · p. 13
Abstract
from arXiv · showhide
Road safety monitoring has historically been reactive, relying on crash-record analysis after fatalities and injuries have already occurred. Proactive identification of high-risk locations and dangerous driving behaviour before incidents occur is a critical but underexplored challenge. This paper addresses this gap using connected vehicle telemetry data from Greater Sydney, Australia, to detect and forecast near-miss risky driving events at the Local Government Area (LGA) level. Risky driving is quantified through g-force thresholds (hard braking >0.6g, harsh cornering >0.47g, harsh acceleration >0.5g), and spatio-temporal heatmaps are constructed to identify high-risk zones. Eight predictive models are benchmarked across three families: ensemble learning (Random Forests, XGBoost, LightGBM), deep learning (LSTM, N-BEATS), and classical time-series methods (ARIMA, Exponential Smoothing, Prophet). ARIMA achieves the lowest mean absolute error (MAE: 162.21), performing comparably to LSTM (MAE: 163.92) and outperforming all ensemble methods, with N-BEATS reaching an MAE of 180.75. These results demonstrate that parsimonious time-series models are competitive with deep learning approaches when training data volume is limited. The study highlights the potential of IoT-based connected vehicle data to support proactive road safety interventions, with Sydney's inner and western LGAs (CBD, Parramatta, Bankstown) identified as persistent high-risk zones warranting targeted policy action.
1. Introduction
The introduction frames Australian road trauma as an urgent public-health problem and motivates proactive near-miss detection using connected-vehicle telemetry. It positions the study as an LGA-level contribution combining validated g-force proxies, predictive-model benchmarking, and actionable risk mapping.
Motivation
- 39,755 individuals were hospitalised in Australia during 2018–2019, including 10,282 who sustained life-threatening injuries.The introduction presents road trauma as a critical public-health issue and notes that 2025 road deaths reached 1,337, the highest level in 12 years.Source: §1.1 · p. 1
Related work
- Event-based in-vehicle data recorders are cost-effective and deployable but may miss incidents not associated with predefined thresholds.Naturalistic driving studies provide continuous, high-fidelity behavioural recording, while connected-vehicle data support spatial clustering of near-crash events.Source: §1.2 · pp. 1–2
Contributions
- The paper presents one of the first Australian LGA-level applications of connected-vehicle IoT telemetry for near-miss prediction beyond retrospective crash analysis.Its contributions include a rigorous data-driven definition, systematic benchmarking across eight models and three families, and practical guidance for limited telemetric data.Source: §1.2 · p. 2
- The study produces place-based risk maps identifying Sydney LGAs prone to dangerous driving to support targeted interventions by transport agencies and local councils.A related Sydney study found statistically significant spatial associations between near-miss clusters and official crash blackspots, validating the g-force proxy.Source: §1.2 · p. 3
2. Case study
The case study uses large-scale connected-vehicle telemetry from Australia to characterize near-miss risky driving by event type, road context, speed, timing, and location. Risky-driving incidents concentrate on urban trunk roads, peak travel periods, and inner and western Sydney LGAs, while near-miss hotspots can differ from historical crash hotspots.
Data and event definition
- Over 700,000 vehicles from 64 manufacturers contributed connected-vehicle data collected from April 2020 to December 2021, after removing the top 1% of g-force events.Data were collected through in-vehicle Data Acquisition Systems with cloud-based transmission.Source: §2 · p. 3
Incident characteristics
- 59.56% of recorded risky-driving actions were braking, 40.13% were steering, and 0.3% were combined events.These classifications were based on manoeuvres recorded in the Compass IoT datasets.Source: §2.2 · p. 4
- 30.3% of incidents occurred on trunk roads, 25.2% on primary roads, and 16.6% on secondary roads; 51.2% affected one lane.Two-lane and three-lane incidents accounted for 22.8% and 19.4%, respectively.Source: §2.2 · p. 4, Figure 3 · p. 5
Temporal and speed patterns
- Most risky-driving events occurred below 20km/h or at medium speeds of 25-40km/h, while events above 100km/h were comparatively infrequent.The study categorized recorded speeds into low-, medium-, and high-impact classes.Source: §2.2 · p. 5
Temporal and spatial patterns
- Weekday peaks occurred at 7-9 AM and 3-6 PM, with Wednesdays and Fridays reaching almost 7,000 events; Sydney’s CBD, Parramatta, and Bankstown ranked among the most dangerous LGAs.Most incidents occurred near cities, while 2,681 Sydney grid cells showed Low-Crash, High-Near-Miss patterns; near-miss and crash counts had Moran’s I of 0.1654 (p < 0.001).Source: §2.2 · pp. 5–6
3. Methodology
The methodology forecasts each LGA’s next-day incident count from its own historical daily risky-driving series, using multiple machine-learning, deep-learning, and time-series models. Models are evaluated with rolling-origin forecasts over a 35-day held-out window using RMSE, MAE, and MAPE.
Forecasting framework
- Each LGA is modelled independently as a univariate daily incident-count time series to support area-level forecasts for targeted resource deployment and preventive measures.The forecasts use historical risky-driving logs to predict the number of incidents likely to occur in each LGA on the following day.Source: §3 · pp. 8–9
Model families
- Eight models span ensemble learning, deep learning, and classical time-series forecasting, including RF, XGBoost, LightGBM, ARIMA, ES, LSTM, Prophet, and N-BEATS.The described models include bagged decision trees, gradient boosting, recurrent neural networks, decomposable forecasting, and basis-function forecasting approaches.Source: §3 · pp. 9–10
Forecasting framework
- All models use only the daily incident-count series per LGA; road geometry, speed limits, weather, infrastructure type, and other external features are excluded from forecasting.For RF, XGBoost, and LightGBM, the univariate series is converted into a supervised learning problem using lagged features.Source: §3 · p. 9
Evaluation protocol
- The evaluation uses an expanding-window rolling-origin procedure over the final 35 days, producing sequential one-step-ahead forecasts at each daily time step.The preceding observations form the initial training set, and model performance is assessed using RMSE, MAE, and MAPE.Source: §3 · p. 9, §3.1 · p. 10
Classical time-series models
- ARIMA uses the default hyperparameters p = 2, d = 1, q = 0 and combines past observations linearly to forecast the next value.ARIMA’s parameters represent autoregression, differencing, and the moving-average component, respectively.Source: §3 · p. 10
4. Results
Across LGAs, ARIMA and LSTM achieved comparable forecasting accuracy, while ensemble models performed poorly. ARIMA’s strong performance is consistent with the task’s simple temporal dependencies and univariate incident-count inputs, though study limitations may explain why complex models underperformed.
Evaluation design
- All models were trained and evaluated independently for each LGA, with MAE, RMSE, and MAPE averaged across LGAs.The models were grouped into ensemble learning, time-series, and neural network classes.Source: §4 · p. 11
Model performance
- MAE: 162.21 for ARIMA was the lowest reported error, while LSTM achieved a comparable MAE: 163.92.Without a formal Diebold–Mariano test, the models should not be definitively ranked.Source: §4 · p. 11
- MAE: 227.19 for XGBoost and MAE: 236.84 for Prophet were associated with systematic over- or under-prediction.These models were identified as the poorest performers across the 35-day evaluation window.Source: Figure 11 · p. 12
Interpretation
- ARIMA suited the task because it contains relatively simple temporal dependencies and linear patterns, while every model used only the univariate incident-count series.Random Forest, XGBoost, and LightGBM received 30-day lagged features from that series.Source: §4 · p. 11
Limitations
- The unexpected underperformance of complex deep-learning models may relate to limitations of the study, which are discussed in the Conclusions section.The passage characterizes traditional models as performing best for the current research problem.Source: §4 · pp. 11–12
5. Conclusions
The study presents a connected-vehicle framework for identifying and forecasting near-miss risky driving, finding concentrated risk in several Sydney LGAs and motivating targeted interventions. Its conclusions also stress pandemic-related limitations, limited deep-learning training data, and the need for real-time extensions and post-pandemic validation.
Conclusions
- Over 700,000 vehicles supported LGA-level identification and short-term forecasting of near-miss risky driving events in New South Wales.Risky driving was defined using g-force thresholds and aggregated geographically at the Local Government Area level.Source: §5 · p. 12
- CBD, Parramatta, and Bankstown were consistently identified as high-risk LGAs, providing a spatial basis for targeted road safety interventions.Recommended actions include prioritising variable message signs and speed enforcement during weekday morning and afternoon peaks.Source: §5 · p. 12
- Post-pandemic data should re-validate spatial priorities before enforcement or infrastructure resources are operationally deployed.The paper also identifies an unaddressed equity dimension associated with risk concentration in Parramatta and Bankstown.Source: §5 · p. 13
Limitations and future studies
- April 2020 to December 2021 data were affected by COVID-19 lockdowns, so reduced incident counts may reflect lower traffic volumes rather than safer driving.Future studies should isolate pandemic-affected periods or apply corrections before drawing policy-relevant conclusions.Source: §5 · p. 13
- Over 700,000 vehicles from 64 manufacturers provided broad geo-spatial coverage, but daily aggregated observations remained insufficient for data-hungry deep-learning architectures.This limitation helps explain why classical time-series models performed competitively.Source: §5 · p. 13
Towards proactive road intelligence
- Ongoing work is developing real-time near-miss prediction from continuously streaming connected-vehicle telemetry and live situational awareness of dangerous driving conditions.This extends beyond the paper’s daily aggregate predictions.Source: §5 · p. 13