Development of pedestrian crash prediction model for a developing country using artificial neural network

Area of Science:

• Transportation Engineering
• Artificial Intelligence
• Road Safety

Background:

• Urban intersections in India experience a high rate of pedestrian fatalities.
• Existing model-based prediction methods for pedestrian fatalities in India are underdeveloped.

Purpose of the Study:

• To develop a predictive model for pedestrian fatal crash frequency at urban intersections in India.
• To explore the effectiveness of Artificial Neural Network (ANN) techniques for this prediction task.

Main Methods:

• Developed 12 Artificial Neural Network (ANN) models using three activation functions and four learning algorithms.
• Optimized ANN models by varying the number of hidden layer neurons through trial and error.
• Identified the best-fit model using Bayesian Regularization Neural Network with a hyperbolic tangent-sigmoid activation function and 13 hidden neurons.

Main Results:

• The Bayesian Regularization Neural Network with 13 neurons and 'hyperbolic tangent-sigmoid' activation was the best-fit model.
• Sensitivity analysis revealed 'approaching speed' of motorized vehicles as the most significant factor influencing fatal pedestrian crashes.
• 'Logarithm of average daily traffic' (ADT) volume was the second most sensitive variable.

Conclusions:

• Artificial Neural Network (ANN) models can effectively predict pedestrian fatal crash frequency at urban intersections.
• Vehicle approaching speed, average daily traffic (ADT) volume, pedestrian-vehicular volume ratio, and pedestrian cross-walk accessibility are critical factors in fatal pedestrian crashes.