Active navigation with reinforcement learning for bus stop amenity auditing in street view imagery

U.S. transit agencies often lack accurate, up-to-date inventories of bus stop amenities (signs, shelters, and seating), which are critical for rider accessibility and service quality. Computer vision applied to street view imagery offers a scalable alternative to labor-intensive field surveys, but existing methods depend on static image retrieval using General Transit Feed Specification (GTFS) coordinates. Because these coordinates are frequently imprecise and amenities are often occluded by vehicles or street furniture, single-viewpoint static retrieval produces unreliable results that existing heuristic-based approaches, such as fixed rotational searches, have failed to adequately resolve. This study develops a reinforcement learning (RL) framework in which a Proximal Policy Optimization (PPO) agent, equipped with a transformer-based feature extractor, actively navigates street view imagery to locate bus stops, acquire multiple viewpoints, and identify amenities. An object detection model provides the reward signal and visual features that guide exploration, while an external spatial memory module records the agent’s path and estimates stop positions. Deployed across 1,919 bus stops in the Atlanta metropolitan area, the agent located stop evidence for 99% of stops, substantially outperforming static retrieval, which captured usable imagery for only one-third of stops. The agent acquired multiple viewpoints of identified stops and triangulated their positions, confirming its capacity to track and assess amenities. This framework offers transit agencies a scalable pathway to acquire system-wide bus stop amenity inventories. More broadly, this work establishes RL-based active navigation as a new approach for urban infrastructure auditing via street view imagery, addressing the limitations of static image retrieval.