A statistical multilevel approach to improve accuracy of Automatic Passenger Counting

Automatic Passenger Counting (APC), or accurate estimation of the number of passengers in public transport, is crucial for optimizing transit operations, improving passenger experience, and supporting data-driven transport planning. Recently, thanks to the increase of popularity of machine learning and AI, two main APC technologies seem to emerge: Wi-Fi-based systems and optical camera systems. The firsts have gained popularity due to their low cost, but they often suffer from accuracy issues caused by signal interference, multiple devices for each person and more importantly the randomisation of addresses. As for camera systems, many commercial APC systems promise very high accuracies, but in real world conditions the accuracy plummets due to poor lighting, bus vibrations, and struggles to count people in crowded situations. Hence, this thesis aims to improve raw Wi-Fi passenger count estimates by performing a comprehensive analysis of influencing factors and applying an innovative statistical modeling technique to enhance accuracy. The methodology provides four steps. The first step is an exploratory analysis of data collected in the cities of Asti and Torino, which contains manual passengers count, also: those were collected by some operators in order to dispose of ground truth data, making it possible to assess Wi-Fi system accuracy. The second step focuses on feature selection and processing: performing correlation analysis allowed to identify the most relevant variables affecting the count. In the third step, the selected features were analysed through a factor analysis, used to group correlated features into meaningful factors that provide a more structured representation of the data. After the deletion of some features having low factor adequacy or low communality, the key finding from this phase is that POIs play a significant role in influencing passenger counts: specific locations (e.g., schools, universities, clothing businesses) seem to attract a higher number of people, and their proximity to bus stops appears to be correlated with bus occupancy levels. Finally, in the last step, in order to verify this association and ensure that it was not incidental (some bus stops may have a large number of POI types, especially in the city centre), a multilevel statistical model was developed. This model seemed particularly suitable for this task, as it accounts for hierarchical data structures, such as random variations across different bus routes, bus stops and hours. By incorporating multiple levels of variability, this approach enables a more robust estimate of passenger count; moreover, the choice to use delta count (obtained as predicted count minus manual count) as dependent variable allowed the multilevel model to assess overestimation and underestimation and to mitigate them. The results show that multilevel model succeeds in halving the mean absolute error and the root mean square error of passengers count: this is a promising result, as it may be possibly used for real-time information systems to display the crowding conditions of the bus, improving the not so accurate Wi-Fi count. In conclusion, this research demonstrates that statistical approaches can complement and enhance traditional sensor-based methods, which are not yet optimal as someone claims. On the other hand, multilevel model requires a certain amount of ground truth data, that is very difficult to find and very costly to produce.