ATOBS: An AI-Based Traffic Optimization and Bus Occupancy Prediction System Using CatBoost

Abstract

Public bus networks in rapidly urbanizing cities exhibit uneven passenger distribution across service routes, with overcrowding during peak hours and underutilization during off-peak periods. Students, first-time commuters and job seekers, who do not know the usual occupancy of the buses do face a lot of inconvenience because of this issue Existing routeplanning systems prioritize travel time and distance, offering no visibility into expected in-vehicle crowding. This paper presents ATOBS (AI-based Traffic Optimization and Bus Occupancy System), a machine learning framework for predicting bus occupancy as a function of route, destination, and time of travel. The system is trained on operational data from the Bangalore Metropolitan Transport Corporation (BMTC). Three regression models-CatBoost, XGBoost, and Random Forest are trained and benchmarked against one another. CatBoost achieves the best performance across all evaluation metrics, with an $\mathrm{R}^{2}$ and Explained Variance Score of 0.81 and the lowest RMSE, MAE, and MAPE among the three models, attributable to its native categorical-feature handling and its capacity to model recurring traffic patterns. By providing occupancy predictions in advance of travel, ATOBS supports informed commuter decision-making and gives transit authorities a datadriven basis for service planning, indicating strong potential for deployment in operational public transport systems.