Development of a sensor fusion methodology for improving performance of an Automated Passenger Counting based on camera and weight in motion system

During the last decade the available data increased exponentially thanks to more affordable electronics and a more powerful computational power, allowing to extract fundamental features or making predictions. During the 2019 pandemics, due to monitoring needs, many public transport operators had to install cameras or other sensors to count the people flow and monitor their temperature. The data coming from those systems can be easily extracted and analysed also for other purposes, such as congestion monitoring, a theme that has a long history in the transport field; it allows both to save money related to personnel and to continuously monitor the state of operation, avoiding the human errors. Lowering the error probability and increasing the accuracy on the counting is fundamental since it allows to properly define the demand, its distribution over the territory and over the time periods of the day, enabling the implementation of fleet adaptive planning and a more efficient scheduling. These last represent a starting point to induce a modal diversion from private vehicles to public transport. To this end, Regione Liguria started the “More Than This” project: different companies were entrusted in the configuration of some connected metro trains. The small fleet mounts different sensors acting as data sources which are sent to a central platform, hosting different mobility services that improve both the operator performances and the user experience. This work focuses on the development of an Automated Passenger Counting systems, in particular, it tries to find a methodology to estimate the number of on-board passengers using the data coming from a “Weight in Motion" system, which was initially installed exclusively for maintenance purposes. Then, after the tuning of the weight-based APC system, the data are fused with the counting measures coming from the vision system. The purpose of the sensor fusion is to lower the uncertainty linked to the two single systems: the cameras have problems with high congestion scenarios while the weight in motion system has problems with the low occupation ones. The two systems together are complementary and the final result is much more accurate than the individual ones. This work not only highlights the power of the sensor fusion, which can be applied only on a small set of equipped vehicles, but it also focuses on providing a counting module for all those trains where the novel platform is not installed. Genova’s metro line has a heterogeneous fleet, so it is not possible to define the onboard passengers using a fixed set of parameters; it is first required to define which type of vehicle is being recorded, so the correct tare information is used for the passenger computation. To this end, an efficient vehicle identification procedure is implemented, able to distinguish the three types of trains by analysing the number of axles sensed and the distance between the first and the last. Finally, after finding the occupation on the various vehicles, a prediction module is studied, trying to find the best method able to forecast the future occupation according to the past values recorded. This is not an easy task since the time series is very similar to a random process, so both machine learning and neural networks performances are analysed.