Source-linked AI summary
Accident Risk Prediction based on Heterogeneous Sparse Data: New Dataset and Insights
Sobhan Moosavi, Mohammad Hossein Samavatian, Srinivasan Parthasarathy, Radu Teodorescu, Rajiv Ramnath
TL;DR
Traffic accident prediction is constrained by small datasets, demanding inputs, and limited real-time applicability. The paper introduces DAP and the US-Accidents dataset using heterogeneous sparse data, and reports improved prediction of rare accident events, especially the accident class.
Problem
Existing accident-prediction studies often rely on small, limited-coverage datasets, extensive data attributes, or setups unsuitable for real-time use.
Method
The paper combines DAP, a deep-neural-network model using traffic, weather, points-of-interest, and time data, with the publicly available US-Accidents dataset.
Results
DAP improves positive-class F1-score by about 16% over the best traditional model and about 7% over the best neural-network-based model.
Takeaways & Limitations
Time, points-of-interest, and traffic data are important for real-time traffic accident prediction, particularly for rare accident events.
Takeaways & Limitations
Direct comparison with state-of-the-art models is not possible because their input types, available data, and reported metrics are inconsistent.
Abstract
from arXiv · showhide
Reducing traffic accidents is an important public safety challenge, therefore, accident analysis and prediction has been a topic of much research over the past few decades. Using small-scale datasets with limited coverage, being dependent on extensive set of data, and being not applicable for real-time purposes are the important shortcomings of the existing studies. To address these challenges, we propose a new solution for real-time traffic accident prediction using easy-to-obtain, but sparse data. Our solution relies on a deep-neural-network model (which we have named DAP, for Deep Accident Prediction); which utilizes a variety of data attributes such as traffic events, weather data, points-of-interest, and time. DAP incorporates multiple components including a recurrent (for time-sensitive data), a fully connected (for time-insensitive data), and a trainable embedding component (to capture spatial heterogeneity). To fill the data gap, we have - through a comprehensive process of data collection, integration, and augmentation - created a large-scale publicly available database of accident information named US-Accidents. By employing the US-Accidents dataset and through an extensive set of experiments across several large cities, we have evaluated our proposal against several baselines. Our analysis and results show significant improvements to predict rare accident events. Further, we have shown the impact of traffic information, time, and points-of-interest data for real-time accident prediction.
1 INTRODUCTION
The paper addresses limitations in accident prediction research by introducing DAP for fine-grained real-time prediction and US-Accidents, a large publicly available dataset built from heterogeneous sparse data. Experiments across large cities show stronger accident-class F1-score than baseline models, while analyses identify important roles for time, points-of-interest, and traffic data.
- Motivation: Existing studies often use small datasets, extensive unavailable attributes, or settings unsuitable for real-time prediction.These limitations include narrow geographic coverage, long prediction intervals, and demanding data prerequisites.
- Approach: DAP combines traffic events, weather, points-of-interest, and time data for prediction within 5km × 5km regions at 15-minute intervals.Its components target heterogeneous temporal, spatial, and contextual information.
- Dataset: US-Accidents contains about 2.25 million accidents from the contiguous United States between February 2016 and March 2019.The dataset includes location, time, descriptions, weather, period-of-day, and points-of-interest attributes.
- Results: DAP improves positive-class F1-score by about 16% over the best traditional model and about 7% over the best neural-network-based model.For combined positive and negative cases, DAP achieves results comparable to the best baselines.
- Contributions: The paper contributes a heterogeneous data-collection methodology, accident insights, and a deep-neural-network solution using sparse data.The proposed system incorporates traffic flow with weather and points-of-interest data for 15-minute prediction.
2 RELATED WORK
Related work studies environmental influences, accident frequency, and binary accident risk. The paper positions its proposal as a real-time risk-prediction approach using shorter intervals, publicly available inputs, traffic events, and points-of-interest data.
- Environmental Stimuli: Prior research examines how weather, traffic flow, and road-network properties relate to accident occurrence or severity.This category includes statistical and data-mining analyses of environmental stimuli.
- Accident Frequency Prediction: Accident-frequency studies predict expected accident counts for road segments or geographic regions using road, weather, imagery, traffic, and historical accident data.Examples use neural networks, convolutional models, LSTM models, and autoencoders.
- Accident Risk Prediction: Accident-risk studies formulate prediction as binary classification for real-time applications, but often use limited geographic coverage or extensive data.Examples include a single highway segment, Tokyo grid cells, and heterogeneous urban data.
- Positioning: The proposal targets accident-risk prediction with 15-minute intervals and uses real-time traffic events and points-of-interest data.The paper also emphasizes inputs that are relatively easy to collect and publicly available.
3 PRELIMINARIES AND PROBLEM STATEMENT
The paper defines traffic events, weather observations, POIs, and geographical regions, then formulates accident-risk prediction over 5km × 5km regions and 15-minute intervals. The task is to predict whether an accident occurs using recent observations while minimizing prediction error.
- Definitions: A traffic event is represented by GPS coordinates, time, categorical type, and natural-language description, with seven event types including accidents, congestion, and construction.
- Definitions: A weather observation records station location and time together with temperature, humidity, pressure, visibility, wind speed, precipitation, and binary weather-event indicators.
- Definitions: A POI is a mapped location represented by latitude, longitude, and type, with annotations adopted from OpenStreetMap definitions.
- Problem formulation: The model operates on square geographical regions of size 5km × 5km and fixed 15-minute time intervals.
- Problem formulation: For each region and interval, the task constructs a representation from traffic events, weather, and POIs, assigns an accident label, and predicts it from the previous eight intervals.
- Problem formulation: The stated objective is to minimize prediction error.
4 ACCIDENT DATASET
The paper introduces US-Accidents, a country-wide traffic accident dataset constructed from an end-to-end data-creation process. It contains 2.25 million U.S. accident cases from February 2016 through March 2019 and is publicly available.
- US-Accidents contains 2.25 million traffic accident cases from the United States between February 2016 and March 2019.
- Figure 1 presents the process used to create the traffic accident dataset.
- The dataset is publicly available through the paper’s listed URL.
4.1 Traffic Data Collection
Traffic data were collected from MapQuest and Microsoft Bing real-time traffic APIs and integrated into a unified dataset. Integration removed likely duplicates using spatial and temporal proximity thresholds.
- The authors collected streaming traffic events from MapQuest Traffic and Microsoft Bing Map Traffic APIs.
- The collection schedule pulled data every 90 seconds from 6am to 11pm and every 150 seconds from 11pm to 6am.
- Records were treated as duplicates when both Haversine distance and occurrence-time differences fell below 250 meters and 10 minutes, respectively.
- About 24,600 duplicated accident records, approximately 1% of all data, were identified under these settings.
4.2 Data Augmentation
The dataset augments traffic records with addresses, weather, POI annotations, and period-of-day labels. POI thresholds are selected by agreement with regular-expression annotations, yielding 30 meters for intersections and 100 meters for junctions.
- Address and weather augmentation: GPS-only accident records were reverse-geocoded with Nominatim into structured address fields.
- Address and weather augmentation: Weather records were paired to each traffic event using the closest station and temporally nearest observation, with an average reporting-time difference of about 15 minutes.
- POI augmentation: The POI process used OpenStreetMap annotations and 27 regular-expression patterns derived from accident descriptions to identify location types such as intersections and junctions.
- POI augmentation: The threshold search selected 30 meters for intersections and 100 meters for junctions because these values maximized annotation correlation.
- POI augmentation: Using the selected thresholds, 27.5% of accident records received at least one available POI annotation.
- Period-of-day augmentation: Accident records were labeled day or night using the TimeAndDate API and four daylight systems.
4.3 US-Accidents Dataset
US-Accidents is a countrywide accident dataset assembled from multiple sources, with analyses describing its temporal, spatial, and road-related characteristics.
- US-Accidents contains about 2.25 million accidents recorded across the contiguous United States from February 2016 to March 2019.
- Time analysis: Accidents occurred more often on weekdays than weekends, with weekday peaks at 8am and 5pm and a single weekend peak at 1pm.
- Points-of-interest analysis: Most accidents occurred near junctions or intersections, while MapQuest reported more near intersections and Bing reported more near junctions.
- Dataset characteristics: The dataset includes map-matching-based road-type coverage and temperature analysis as part of its documented characteristics.
5 ACCIDENT PREDICTION MODEL
The DAP framework represents each region over 15-minute intervals using heterogeneous traffic, time, weather, POI, and spatial features, then combines temporal and spatial processing for accident prediction. Its architecture uses recurrent modeling for time-variant inputs, POI processing for spatial characteristics, and fully connected layers for final prediction.
- Feature representation: DAP represents each 5km × 5km region during 15-minute intervals with traffic, time, weather, POI, and other spatial attributes.The representation contains 24 time-variant and 113 time-invariant attributes, with the last eight intervals covering two hours.
- Feature representation: The feature vector includes seven traffic-event frequencies, 10 weather attributes, and 13 POI frequencies for spatial characteristics.The traffic categories include accidents, broken vehicles, congestion, construction, events, lane blocks, and flow incidents; weather also includes rain, snow, fog, and hail indicators.
- DAP architecture: DAP processes eight 24-feature time-variant vectors as a temporal sequence through an LSTM component with two layers of 128 cells.The recurrent component outputs a vector of size 128.
- DAP architecture: The framework combines heterogeneous data sources through recurrent, POI, and fully connected components in the Deep Accident Prediction model.Figure 4 identifies DAP as a deep neural-network-based accident prediction model and defines its recurrent and fully connected framework components.
- DAP architecture: A POI component feeds the 13-dimensional spatial vector into a 128-unit sigmoid feed-forward layer, while the fully connected component produces the final prediction.The fully connected component uses dense layers of sizes 512, 256, 64, and 2.
6 EXPERIMENTS AND RESULTS
Experiments evaluate DAP against traditional and neural-network baselines across six cities using F1-scores for accident, non-accident, and weighted-average classes. DAP improves accident-class prediction, while traffic, POI, and time-related features contribute to performance.
- Experimental setup: Experiments use six cities, 12 weeks of sampled data, and separate training and test periods.The selected cities represent diverse traffic, weather, population, density, and urban characteristics.
- Models and metrics: The evaluation compares DAP with logistic regression, gradient boosting, DNN, and DAP-NoEmbed using class-specific and weighted-average F1-scores.Each model was run three times, with average results reported.
- Model comparison: DAP provides a significant improvement for the accident class, whereas LR and GBC perform slightly better for the non-accident class.For weighted-average F1-score, LR, DAP, and GBC slightly outperform the other models; accident-class performance receives greater emphasis because accidents are rare.
- Comparison limitations: Direct comparison with state-of-the-art models is not possible because of inconsistent input types, unavailable data, and inconsistent reported metrics.The paper notes that prior F1-score results show similar trends and values.
- Feature analysis: Removing Desc2Vec, POI, and traffic attributes together causes a significant performance drop, although individual removals may have limited effects because categories can be correlated.The reported feature-removal analysis uses GBC and DAP-NoEmbed.
- Feature analysis: Among POI types, crossing, junction, stop, and traffic signal are more effective for accident prediction than the other examined types.
7 CONCLUSION AND FUTURE WORK
The paper addresses limitations in prior real-time accident-prediction studies with a sparse-data framework and the publicly available US-Accidents dataset. Experiments report improved rare-accident prediction and identify influential data categories.
- Conclusion: The framework targets real-time traffic accident prediction using easy-to-obtain sparse data and neural-network components for traffic events, weather, POI, and time.
- Conclusion: US-Accidents is a publicly available countrywide dataset created through data collection, cleansing, and augmentation.
- Conclusion: Extensive experiments compare the framework with neural-network-based and traditional machine-learning models and show its superiority.
- Conclusion: Time, traffic events, and points of interest are identified as having an impact on traffic accident prediction.