Automatic passenger counting (APC) systems are crucial as they provide accurate data for planning and operations, enable demand analysis and route optimization, contribute to fare collection and revenue management and assist in performance monitoring and service quality evaluation. We have recently developed an automatic passenger counting system based on a neural network (NAPC) that is using data from special 3D counting sensors (APC sensors). These are mounted above the door area and view the boarding area from above. Such APC sensors are costly to purchase and maintain and are therefore typically not installed in every vehicle, but only in a part of the total fleet (e.g. 30% at BVG). In order to achieve a higher equipment rate in vehicle fleets, we aim to develop a new NAPC using surveillance cameras such as CCTV that are already installed for security purposes in most public transportation systems. As a part of our project, the task is to develop an NAPC based on 2D video data, where a U-Net or Feature Pyramid Networks is used for extracting important information, such as depth information.
Requirement: good programing skills in Pytorch or tensorflow