Passenger Counting in Mass Public Transport Systems using Computer Vision and Deep Learning

Abstract

Estimating the number of people in vehicles and stations in public transport systems is crucial to improve service quality. The TransMilenio system in Bogota has serious drawbacks due to the lack of information in congestion situations. In this work we present a computer vision method that estimates the number of people in TransMilenio stations using deep learning techniques. We offer free use of the TransMilenio-Javeriana database with nearly 900,000 head labels on buses and stations. From these images a deep learning architecture tuned for crowd counting was trained to generate density maps around the heads in the scene. Several head count methods were evaluated on the density maps. After testing the method with 10,800 images, the results show a mean absolute error of 1 head per frame, equivalent to 11% relative error. The accuracy of this method is much better than its manual counterpart. This method is also scalable and low cost, which indicates that it has great potential to provide information for the planning and operation of public transport systems.