Indian Pedestrian Behaviour Modelling Using Imitation and Reinforcement Learning

Pedestrian modelling is a critical element in urban traffic simulations, particularly in environments with high variability and unpredictability, such as Indian cities. With the rise of machine learning, advanced techniques now enable more realistic and dynamic representations of pedestrian behaviour. These models play a key role in urban planning, traffic system design and the development of autonomous navigation systems. The current research presents a novel modeling technique that represents pedestrian behavior within Indian urban traffic environments. The technique employs both reinforcement learning (RL) and imitation learning (IL) to animate pedestrian movements that resemble a humanoid character. The agent adopts multiple models to enable navigation in the urban environment. While Policy Optimization Algorithm helps the agent navigate a road crossing scenario, Behavior Cloning imitates the natural walking style of a pedestrian. The model assists in replicating various urban encounters, including active marketplaces, neighbourhoods, and busy traffic crossings — all characteristic of Indian cities. Pedestrian behavior exhibits variability across different scales, ranging from individual decision-making to collective dynamics in large crowds. The model and the scenarios were validated for their robustness through Monte Carlo simulations. The success rates observed ranged from 33% in high-speed traffic conditions to 88% in moderate-speed environments. During training, the agents attained an average reward of 0.981 with a normalized mean. The realism of the imitation learning (IL)-generated models was validated through a Turing test, in which 59.7% of participants misclassified the movements of AI-generated pedestrians as human.