New AI System Uses Bus Cameras to Track Passenger Flow Without Tracking Individuals

Public transit agencies struggle to collect accurate, cost-effective data about where passengers board and exit buses—information essential for route optimization and service planning. Current methods rely on manual surveys, Bluetooth tracking, or passenger counters, all of which are expensive or lack individual-level precision. A new research framework called TransitReID offers an alternative: leveraging existing onboard surveillance cameras to automatically track passenger movement patterns while processing all data locally to preserve privacy.

The Problem with Current Data Collection#

Transit operators need Origin-Destination (OD) data to understand passenger flows and improve service efficiency. Existing approaches have significant limitations: manual surveys are labor-intensive, Bluetooth and WiFi tracking depend on passenger devices, and automated counters cannot match individuals across boarding and exit points. Yet buses and transit vehicles already carry cameras—infrastructure that remains underutilized for this purpose.

How TransitReID Works#

The framework tackles the technical challenges unique to transit environments through three main innovations:

• Occlusion-resistant identification: An algorithm combining a variational autoencoder with region-attention mechanisms learns to recognize passengers even when partially obscured by other riders or luggage. Rather than analyzing full bodies, the system emphasizes visible and discriminative body regions.

• Dynamic matching for real operations: A hierarchical storage and matching system adapts to the continuous flow of passengers boarding and exiting, balancing accuracy against the computational demands of live processing.

• Edge-based privacy protection: The system runs on NVIDIA Jetson devices deployed directly on vehicles, meaning passenger images never leave the bus—only aggregated movement data is transmitted.

Performance and Real-World Testing#

The researchers built a new dataset containing over 17,000 images from actual bus front and rear cameras, captured under varying occlusion and viewing angles. On this transit-focused benchmark, TransitReID achieved 88.3% accuracy in identifying individuals across camera views. In both simulations and real bus operations, the system maintained 80–90% accuracy in estimating OD patterns while operating at near real-time speeds on edge hardware.