Description

This project focuses on understanding and modeling public transport bus driver behaviors under heterogeneous traffic conditions using GPS trajectory data, speed, weather, topology, and metadata. While some research exists for drivers in homogeneous traffic settings, there is a significant gap in analyzing driver behavior under more complex, mixed traffic environments. By applying machine learning techniques, the project aims to identify and cluster strategic and aberrant driver behaviors, contributing to safer and more efficient public transit systems. The outcomes will support transit authorities, urban planners, and policy makers in promoting sustainable and smart mobility.

Motivation and Research Objective

Motivation

• Strategic Advantage: Improve passenger comfort and service quality in competitive transport sectors.

• Road Safety: Detect risky behaviors early to reduce accidents and enhance safety.

• Autonomous Vehicle Insight: Provide critical training data for behavior modeling in self-driving systems.

• Insurance & Policy: Enable fair premium calculations based on real behavior profiles.

Research Objectives

• Identify and analyze driver behaviors in heterogeneous traffic using GPS and related data.

• Develop clustering models to group driver behavior patterns.

• Create predictive models for future driver behavior based on historical trajectories.

• Build a real-time dashboard for monitoring and forecasting driver behavior patterns in public transport.

Results and Impact

Our study delivered a comprehensive analysis of public transport driver behavior under heterogeneous traffic using GPS data, speed, dwell time, and topological features. The results are categorized into three main areas: Analysis, Clustering, and Impact.

Behavioral Analysis

• At bends, drivers were grouped as aggressive, normal, or cautious, based on speed trends.
  A noticeable drop in speed from 1–2 PM across all drivers suggests adaptation to school traffic.

• Speed zone analysis showed that drivers naturally slowed in urban areas like Kandy and sped up in suburban areas like Digana, especially in less complex zones.

• Dwell time varied by time and direction:

  • Lower in the morning from Digana to Kandy

  • Higher in the evening in the same direction, due to shifting passenger demand
    Drivers were classified into four dwell behavior groups: Low, Average, High, and Adaptive dwellers.

Driver Clustering

• Using K-means clustering, we segmented driver behavior into aggressive, normal, and calm groups based on variables like speed, acceleration, and variance.

• 10-minute interval segmentation produced the most meaningful clusters, verified using silhouette scores and Davies-Bouldin Index.

Impact

• Enables real-time monitoring and forecasting of driver behavior via dashboards.

• Supports transport safety policies, driver coaching, and autonomous vehicle insights.

• Lays a foundation for future integration with insurance risk profiling and smart city mobility systems.