EQUIFLOW: A Heterogeneous Graph based Deep Reinforcement Learning Framework for Adaptive Traffic Signal Control at Indian Urban Intersections

Most adaptive traffic signal management research centers around optimizing vehicle speed, and considers passengers, vehicles, and emergency vehicles as an afterthought. Almost none of this study was evaluated toward the kind of mixed and lane-discipline-less traffic observed on Indian roads. In this paper we suggest an equity-sensitive multimodal traffic control system termed EquiFlow, built and tested on the actual 4-junction corridor on ISB Road in Hyderabad. EquiFlow encodes vehicles, lanes, signal phases, pedestrian crossings, and bus stops as separate node types in a heterogeneous transportation graph, which is then encoded by a shared heterogeneous Graph Transformer-based encoder and input to a single proximity-based Policy Optimization (OP) agent, with one distinct action head per typical intersection. A five term reward system penalizes a combined vehicle waiting period, length of queue, a pedestrian delay, bus schedule variance pulled in actual TGSRTC GTFS data, and emergency vehicle delay, with an automated ranking boost when ambulance or equivalent vehicle is identified. The simulation in itself does not involve guesswork but is based on actual data. It employs YOLO based vehicle and pedestrian identification to validate the composition of traffic and validates it with Indian Roads and Congress capacity based guidelines. The resultant policy clearly surpasses the Fixed Time and max pressure baselines, as well as a graph-free MLP PPO treatment based on the demand pattern it had been trained on. On a demand pattern that it has not encountered earlier, each controller falls roughly at the same level, that says something significant: a market demand variance, not the controller itself.