💻 computer science

Predicting Traffic Accidents and Fatalities with Transformer-Based Models and Big Data for Enhanced Policy and Safety Insights

This study proposes a scalable, big data-driven framework utilizing a Spatio-Temporal Parallel Transformer (STPT) and clustering techniques to accurately predict traffic accidents and fatalities in Bangkok, thereby enabling evidence-based policy interventions to enhance urban road safety.

Original authors: Sudarat Sukjaroen, Xiaodan Dong, ST Boris Choy, Weidong Huang, Adel Fadhl Noor Ahmed

Published 2026-07-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sudarat Sukjaroen, Xiaodan Dong, ST Boris Choy, Weidong Huang, Adel Fadhl Noor Ahmed

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine Bangkok's roads as a giant, chaotic game of tag played by millions of cars, motorcycles, and pedestrians every single day. Sometimes, the game goes perfectly; other times, players crash into each other, and sadly, some don't get back up. A team of researchers decided to stop playing "guess who" and instead built a super-smart digital detective to figure out exactly when and where these crashes are most likely to happen, and how bad they might get.

The Detective's Toolkit: A Crystal Ball Made of Math
Instead of just looking at a map and guessing, the team fed their detective a massive buffet of data: 10,163 accident records from 2019 to 2024, traffic counts from 387 different spots, weather reports, and even the specific types of vehicles involved. They didn't just use one brain; they built a whole team of digital brains.

The star player was a new kind of AI called a Transformer (specifically, a Spatio-Temporal Parallel Transformer, or STPT). Think of this model like a super-organized librarian who can read every book in the library at the exact same time, noticing how a story in one section connects to a story in another section miles away. It looked at both time (when the crash happened) and space (where it happened) all at once.

The Big Reveal: What the Data Actually Says
The results were pretty cool, but they came with a few important caveats.

The Best Detective: The Transformer model was the best at predicting the number of accidents and fatalities, making the fewest mistakes compared to older methods like standard math equations or simpler AI. However, the paper notes that because fatal accidents are rare (like finding a needle in a haystack), the model sometimes underestimated the number of deaths. It's great at saying "nothing bad will happen," but a little shy about predicting the really big, rare disasters.
The Runner-Up: A model called Gradient Boosting was also a strong contender. It was really good at distinguishing between a minor fender-bender and a serious crash, acting like a sharp-eyed referee.
What They Ruled Out: The study explicitly showed that older, simpler models (like basic Linear Regression) weren't as good at spotting the complex patterns in this messy data. They also found that just looking at traffic volume wasn't enough; you needed to mix in where the road was, what the weather was like, and who was driving.

The "Hot Zones": Where the Danger Hides
By grouping the city into 10 different "traffic density clusters" (like sorting a messy room into 10 different piles), the researchers found some surprising patterns:

The "Low Traffic" Trap: You might think busy roads are the most dangerous, but the second-highest accident rate happened on quiet, unknown roads with very little traffic. Here, the main culprit was speeding. Drivers on these straight, flat roads seemed to think, "No one is watching, so I'll zoom!" This led to lots of rear-end collisions, especially involving motorcycles.
The "Cargo Truck" Zone: Another hotspot was a specific route where cargo pickup trucks were the main troublemakers. The cause? Running red lights and ignoring traffic signs on straight roads.
The "Fatality" Factor: When it came to the most deadly spots, the pattern shifted. The highest predicted fatality rates happened on rural roads involving motorcycles and speeding. The second-highest was in the city, but this time, it was all about pedestrians getting hit, often due to drunk driving or people cutting off cars.

The "What If" Game (And What It Isn't)
The researchers used a "Decision Tree" (which is like a flowchart of "If this, then that") to figure out exactly what conditions led to these crashes. For example, they found that if a road is straight, has no slope, and a motorcycle is involved, the accident risk jumps up.

The Verdict: A Map, Not a Magic Wand
The paper suggests that this new framework is a powerful tool for city planners. It's not a magic wand that stops all crashes instantly. Instead, it's a high-tech map that says, "Hey, if you put speed cameras on these specific straight roads, or if you teach motorcyclists to wear better gear in these specific rural areas, you might save lives."

The authors are careful to say that while their model is better than the old ones, it's still learning. They suggest that in the future, if they could add video footage of drivers' faces and behaviors, the detective might get even smarter. For now, though, they've proven that mixing big data with fancy AI can help us see the invisible patterns that lead to accidents, giving us a better chance to fix them before they happen.

Technical Summary: Predicting Traffic Accidents and Fatalities with Transformer-Based Models and Big Data

Problem Statement
Road traffic safety remains a critical public health challenge, particularly in metropolitan areas like Bangkok, Thailand, which experiences high accident rates and fatalities. While deep learning has shown promise in accident prediction, existing models often focus on specific accident types (e.g., rear-end conflicts) or fail to adequately integrate complex spatial and temporal heterogeneity. There is a need for a scalable framework that jointly predicts accident frequency and fatality counts by incorporating a wide range of variables—including traffic volume, environmental conditions, driver behavior, and vehicle composition—to provide evidence-based intelligence for targeted policy interventions.

Methodology
The study proposes a comprehensive framework integrating big data analytics with advanced deep learning architectures. The methodology involves three primary stages:

Data Integration and Preprocessing:
- Datasets: The study combines a government dataset of 9,729 road traffic accident records (January 2019–July 2024) with traffic volume data from 387 monitoring stations.
- Features: The predictor vector ( $X_i$ ) includes 37 dimensions categorized into temporal attributes, spatial/route information, vehicle/human involvement measures, and casualty-related variables.
- Clustering: Unsupervised K-means clustering is applied to traffic volume data to identify 10 distinct traffic-density clusters. This step captures spatial heterogeneity and allows for the analysis of accident patterns under varying traffic conditions.
Modeling Approaches:
The study evaluates nine models, with a specific focus on two novel deep learning architectures:
- Spatio-Temporal Parallel Transformer (STPT): This architecture integrates Transformer encoders and Graph Convolutional Networks (GCNs) in parallel. It utilizes a self-attention mechanism to capture long-range temporal dependencies and graph structures to model spatial correlations between nodes (e.g., latitude/longitude proximity). The output is a weighted fusion of spatial and temporal features controlled by a learnable coefficient $\alpha$ .
- Graph Neural Network (GNN): This model represents road segments as nodes and learns relationships via message passing on a graph defined by geographic proximity or traffic similarity. It aggregates information from neighboring nodes to model localized spatial dependencies.
- Baseline Models: For comparison, the study includes Gradient Boosting, Decision Trees, XGBoost, LightGBM, Random Forest, Linear Regression, Support Vector Regression (SVR), and standard Neural Networks.
Analysis and Interpretation:
- Model performance is evaluated using Mean Squared Error (MSE), Mean Absolute Error (MAE), and Area Under the Receiver Operating Characteristic Curve (AUC-ROC).
- Decision Tree analysis is employed on the clustered data to identify specific conditions (e.g., traffic volume, vehicle type, road geometry) leading to high accident and fatality rates.
- Geospatial hotspot mapping is used to visualize high-risk corridors.

Key Contributions

Novel Architecture Application: The study fills a research gap by applying the Spatio-Temporal Parallel Transformer (STPT) to predict both accident occurrence and fatality rates, moving beyond single-type accident prediction.
Comprehensive Variable Integration: Unlike previous studies, this framework jointly analyzes 37 variables across spatial, temporal, behavioral, and environmental dimensions.
Cluster-Based Insight: By integrating K-means clustering with predictive modeling, the study reveals how accident patterns differ significantly across traffic density levels, identifying specific clusters with abnormally high accident or fatality rates.
Actionable Policy Framework: The research translates model outputs into specific, targeted mitigation strategies for identified hotspots, such as speed enforcement on straight roads or pedestrian safety measures.

Results

Predictive Performance: The STPT model achieved the lowest Mean Squared Error (MSE = 0.0009), outperforming all other models in predicting fatality rates. However, the authors note that the model underestimated fatalities in rare, high-severity events due to class imbalance in the dataset.
Classification Capability: The Gradient Boosting Regressor demonstrated the highest discriminative power with an AUC of 0.8405, followed by the GNN (AUC = 0.8103). The STPT model showed moderate classification performance (AUC = 0.7783).
Pattern Discovery:
- Temporal Trends: Accident rates peak during weekday rush hours (8–9 AM, 4–5 PM) and holiday periods (December/January).
- Cluster Analysis: Clusters 5 and 6 exhibited the highest accident rates, while Cluster 9 showed the highest fatality rate.
- Hotspot Specifics:
  - Second-Highest Accident Rate: Associated with low traffic volume, unknown routes, motorcycles, speeding, and rear-end collisions on straight roads.
  - Highest Fatality Rate: Linked to rural roads, motorcycles, speeding, and straight roads with slopes.
  - Pedestrian Fatalities: Correlated with drunk driving, negligence, and collisions involving cargo pickup trucks on straight roads.

Significance and Claims
The paper claims that the proposed framework offers a transferable and scalable approach for urban road safety assessment. By leveraging big data and deep learning, the study demonstrates the ability to identify high-risk conditions and hidden patterns that traditional statistical methods might miss. The authors assert that this evidence-based intelligence enables targeted interventions—such as infrastructure improvements, speed enforcement, and public awareness campaigns—ultimately contributing to the reduction of traffic-related injuries and fatalities. The study emphasizes that while the models show high predictive accuracy, future work should address class imbalance and integrate multimodal data (e.g., visual streams) to further enhance proactive safety measures.

Technical Summary: Predicting Traffic Accidents and Fatalities with Transformer-Based Models and Big Data

More like this