Real-time traffic accidents post-impact prediction: Based on crowdsourcing data.

Area of Science:

• Intelligent Transportation Systems
• Data Science
• Traffic Engineering

Background:

• Effective traffic accident management is crucial for intelligent transportation systems.
• Crowdsourcing and floating car data offer new avenues for improving traffic management.
• Predicting traffic flow evolution post-accident is complex but vital.

Purpose of the Study:

• To investigate methods for predicting traffic flow evolution after accidents using crowdsourcing data.
• To develop a hierarchical scheme for identifying congestion levels and predicting their duration.
• To evaluate the effectiveness of machine learning algorithms in this prediction task.

Main Methods:

• Categorizing traffic conditions into four levels (severely congested, congested, slow moving, uncongested) based on congestion delay index.
• Defining four accident types corresponding to these congestion levels.
• Implementing a hierarchical scheme to identify the most congested level and predict duration.
• Validating the model using 2017 Beijing traffic accident data with Random Forest (RF), Support Vector Machine (SVM), and Neural Network (NN) algorithms.

Main Results:

• Neural Network (NN) demonstrated superior performance in absolute difference assessments.
• Random Forest (RF) slightly outperformed Support Vector Machine (SVM), particularly for short-term, severely congested predictions.
• Continuous updating of traffic condition information significantly improved prediction accuracy across all models.

Conclusions:

• Crowdsourcing data is effective for real-time traffic accident analysis.
• The proposed hierarchical model is a viable tool for analyzing traffic data post-accidents.
• Machine learning models, especially NN and RF, can enhance traffic accident management strategies.