BDRoadRisk: Large-Scale GPS Traffic Accident Severity Dataset
Description
Dataset Overview: This dataset contains 10 million synthetic GPS-based traffic event records simulating road conditions on the urban road networks of Dhaka and Chittagong, Bangladesh. Each record includes 23 features and is labelled with one of three accident severity classes: No_Accident, Minor, and Major. The features are designed to include both pre-accident and post-accident information, where pre-accident features describe environmental, road, and vehicle conditions before an accident, while post-accident features capture factors associated with the accident event and its severity. Severity labels are generated using a context-driven probabilistic model. A 1.8% boundary label noise is applied at the No_Accident/Minor boundary to reflect real-world reporting ambiguity. The dataset is provided in two formats: a Hive-partitioned Apache Parquet archive compatible with Apache Spark and distributed query engines, and a single flat CSV file suitable for Pandas and R. It is intended for multi-class severity classification benchmarking, explainability research, and big data pipeline development in intelligent transportation systems. The attached data dictionary provides detailed descriptions of all features, including their categorization into pre-incident and post-incident groups. Dataset Composition: - Total Records: 10 million - Total Features: 23 - Target Class Labels: No_accident, Minor and Major. - Class Distribution: No_accident 74.41%, Minor 17.30%, and Major 8.28%. Features (Attributes): - latitude (Continuous decimal) - longitude (Continuous decimal) - speed_kmh (Continuous, road-class dependent) - acceleration_ms2 (Continuous) - visibility_km (Continuous) - temperature_celsius (Continuous, non-predictive) - traffic_density (Integer, non-monotonic) - vehicle_age_years (Integer, right-skewed) - driver_age (Integer, weak signal) - reaction_time_sec (Continuous, time-conditional) - weather_condition (Clear, Rain, Fog, Storm) - road_surface (Dry, Wet, Icy, Damaged) - junction_type (None, T-Junction, Crossroad, Roundabout) - light_condition (Daylight, Dusk_Dawn, Dark_Lit, Dark_Unlit) - time_of_day (Morning, Afternoon, Evening, Night) - vehicle_type (Car, Truck, Motorcycle, Bus, Van) - driver_experience_level (Novice < Intermediate < Experienced) - road_quality (Poor < Fair < Good) - speed_limit_zone (30 < 50 < 80 < 100 < 120 km/h) - alcohol_influence (None < Suspected < Confirmed) - seatbelt_worn (Yes > No) - month (1–12) - road_class (motorway, primary, secondary, residential) - accident_severity (No_Accident, Minor, Major) Research Implications: The dataset supports the development and validation of ML and DL models for post-incident traffic accident severity classification. Additionally, the dataset can be utilized for accident incident classification (No Accident vs. Accident) as well as accident severity classification (Minor vs. Major).
Files
Steps to reproduce
The dataset is generated using a Python-based stochastic simulation engine, designed to create 10 million records. 1. Environment and Setup: The simulation is implemented in Python 3.12.13 using NumPy 2.0.2 for probabilistic sampling and Pandas 2.2.2 for data assembly. A fixed random seed (42) ensures reproducibility. The dataset is generated in partitions defined by road_class and month. 2. Spatial and Temporal Partitioning: Each record is assigned to either Dhaka/Chittagong with fixed latitude/longitude ranges. Data is generated separately for each combination of road_class and month, ensuring seasonal variation across 48 partitions. 3. Context Feature Generation (Pre-accident conditions): For each partition, environmental and structural features are sampled first, independent of accident outcome: - Weather is sampled using season-aware probabilities . - Road surface and light conditions are derived from weather and time_of_day. - Vehicle type, road quality, junction type, and speed limit are generated using road-class-specific probability distributions. - These variables define the contextual driving environment before any accident occurs. 4. Risk Modeling (Accident probability construction): A context risk score is computed by combining weighted effects of road class, weather, time of day, vehicle type, road quality, surface condition, junction type, and seasonality. Non-linear interaction terms further amplify or reduce risk. The final risk score is clipped to maintain realistic bounds. 5. Severity Assignment (Probabilistic labeling): Accident severity is sampled per record using the computed risk score: - Base accident probability is scaled by context risk. - Conditional probability determines class distribution across No_Accident, Minor, and Major. - Final global distribution is calibrated to approximately 75% / 17% / 8%. - A 1.8% label noise is applied between No_Accident and Minor classes. 6. Post-accident Feature Generation: After severity assignment, dynamic features are generated as a function of risk and class: - Speed, acceleration, visibility, traffic density, and reaction time are sampled using severity-conditioned distributions. - Higher severity leads to higher speed variance, stronger braking, reduced visibility, higher congestion, and slower reaction time. - Minor class includes additional heteroscedastic noise to simulate real-world ambiguity. 7. Categorical Post-accident Adjustments: Certain behavioral and event-based features are sampled conditionally on severity: - Driver experience level, alcohol influence, and seatbelt usage depend on accident outcome. - Junction type is a 50/50 blend of road-class-based and severity-based distributions. 8. Output and Storage: Each partition is split into 100 chunks and saved in Apache Parquet format with Snappy compression for efficient distributed processing. The full dataset contains 10 million rows and is organized by: road_class=... / month=...
Institutions
- Chittagong University of Engineering & TechnologyChittagong, Chittagong