SMART AUTONOMOUS VEHICLE PERCEPTION USING REINFORCEMENT LEARNING–DRIVEN V2I SENSOR FUSION

Abstract

Autonomous driving research has evolved from rule-based vehicle systems to data-driven perception using V2I communication and multi-sensor fusion. In India, rapid urbanization has led to severe traffic congestion, with over 1.5 lakh road fatalities annually and average urban traffic speeds below 25 km/h. Government initiatives like Smart Cities Mission and Intelligent Transportation Systems (ITS) emphasize V2I-enabled safety, making AI-driven perception and traffic intelligence highly relevant. This framework aims to enhance autonomous vehicle perception by fusing infrastructure and vehicle sensor data using machine learning and reinforcement learning to accurately predict traffic density and vehicle priority levels. In traditional systems, traffic monitoring relies on fixed rule-based logic, human-operated traffic signals, roadside cameras, and manual control centers. Traffic density is estimated through periodic surveys or static sensors, while vehicle priority is assigned using predefined rules for emergency or public vehicles without adaptive intelligence. Manual systems lack realtime adaptability, are prone to human error, and cannot effectively handle dynamic traffic patterns. They fail to scale with increasing vehicle density, provide delayed responses, and are unable to learn from historical data, leading to inefficient traffic flow and reduced road safety. The motivation arises from limitations of manual and rule-based systems in handling complex, dynamic traffic environments. This research aims to overcome poor scalability, delayed decision-making, and lack of learning capability by introducing intelligent, data-driven models that adapt in real time and improve perception accuracy. The proposed V2I-MSF framework integrates multi-sensor data from vehicles and roadside infrastructure and applies machine learning and reinforcement learning for intelligent perception and decisionmaking. SVM and Random Forest models classify vehicle priority and predict traffic density, while stacked RF-HST-LLR ensembles improve robustness and accuracy.