CAF – Passenger Counter Based on Computer Vision
CAF
Sector: Mobility
Business Case
We propose to develop a passenger counter based on computer vision. There is currently one infrared sensor per door for passenger counting. In general, each functionality needs independent sensors. We are looking for a comprehensive solution that combines different functionalities in one single solution.
Objectives
The objective of the use case is twofold: on the one hand, to develop a vision-based passenger counter that improves the current infrared-based counting system, and on the other hand, to generate a single scalable intelligent system capable of integrating different functionalities.
Use case
On the one hand, we propose to define the set-up of cameras and on-board computers in an optimal and feasible way. On the other hand, we propose to develop a solution (based on microservices) that is architecture agnostic and easily scalable. The deployment is to be carried out on board the train.
Infrastructure
Edge On Premise Cloud
Technology
Machine learning and deep learning Computer vision
Data
Los datasets son privados. Contienen imágenes generadas por cámaras de ojo de pez con dato sensible (rostros de personas). Es necesario considerar la ley de protección de datos.
Resources
At the organisational level, 1 project manager, 1 HW technician (cameras and computing) and 2 SW technicians (computer vision and microservices) were necessary. Technical development mainly by Vicomtech, and technical validation, requirements and feasibility by CAF.
Difficulties and learning
La gran dificultad que hemos tenido ha sido cumplir con la ley de protección de datos. En problemas de visión el gran cuello de botella es el etiquetado. Trabajar con microservicios hace que la solución sea agnóstica de arquitectura y fácilmente escalable. Esta tecnología es válida siempre y cuando no necesitemos baja latencia. No es necesario generar las inferencias para cada frame.
KPIs (business impact and metrics of the model)
mAP (mean Average Precision): In this case, having only one class (person) is equivalent to AP (Detection), accuracy (Detection), recall (Detection), counting accuracy (Tracking), frame rate of inferences (Perception), message processing frequency (Kafka).
Funding
The project was funded by the Guipuzcoan Network of Science, Technology and Innovation programme.
Collaborators, Partners
Vicomtech
