Large Scale Passenger Detection with Smartphone/Bus Implicit Interaction and Multisensory Unsupervised Cause-effect Learning

TL;DR

This work tackles scalable Be-In/Be-Out (BIBO) detection essential to MaaS by fusing GPS data from buses with BLE beacon signals from the infrastructure, avoiding reliance on manual high-quality labels. It introduces the Cause-Effect Multitask Wasserstein Autoencoder (CEMWA), a dual-encoder architecture that learns a 4D latent representation capturing the cause–effect relationship between smartphone proximity and BLE signal strength ($X=(X_1,X_2)$ with $X_1$ as GPS and $X_2$ as RSSI). By combining multi-task loss weighting (via Kendall's uncertainty), a regularized WAE-MMD objective using kernels like $k^{\text{RBF}}$, and a deep clustering step with DBSCAN, CEMWA yields an unsupervised BI/BO classifier that resists label noise and supports scalable, privacy-conscious sensing for MaaS. Empirical results on a dataset from three autonomous buses in Denmark show competitive performance and resilience to noisy ground truth, highlighting the method’s potential for ticketless transit systems and cross-operator data sharing within MaaS. The approach promises energy-efficient, scalable passenger detection with broad applicability to real-world transport networks and related implicit sensing tasks.

Abstract

Intelligent Transportation Systems (ITS) underpin the concept of Mobility as a Service (MaaS), which requires universal and seamless users' access across multiple public and private transportation systems while allowing operators' proportional revenue sharing. Current user sensing technologies such as Walk-in/Walk-out (WIWO) and Check-in/Check-out (CICO) have limited scalability for large-scale deployments. These limitations prevent ITS from supporting analysis, optimization, calculation of revenue sharing, and control of MaaS comfort, safety, and efficiency. We focus on the concept of implicit Be-in/Be-out (BIBO) smartphone-sensing and classification. To close the gap and enhance smartphones towards MaaS, we developed a proprietary smartphone-sensing platform collecting contemporary Bluetooth Low Energy (BLE) signals from BLE devices installed on buses and Global Positioning System (GPS) locations of both buses and smartphones. To enable the training of a model based on GPS features against the BLE pseudo-label, we propose the Cause-Effect Multitask Wasserstein Autoencoder (CEMWA). CEMWA combines and extends several frameworks around Wasserstein autoencoders and neural networks. As a dimensionality reduction tool, CEMWA obtains an auto-validated representation of a latent space describing users' smartphones within the transport system. This representation allows BIBO clustering via DBSCAN. We perform an ablation study of CEMWA's alternative architectures and benchmark against the best available supervised methods. We analyze performance's sensitivity to label quality. Under the naïve assumption of accurate ground truth, XGBoost outperforms CEMWA. Although XGBoost and Random Forest prove to be tolerant to label noise, CEMWA is agnostic to label noise by design and provides the best performance with an 88\% F1 score.

Figures (13)

• Figure 1: Cause-effect Multi-task Wasserstein Auto-encoder (CEMWA) independent cross-reconstruction of $X_1, X_2$ minimizing \ref{['eq:CEMWA']} and clustering of the resulting latent space, 5028 parameters.
• Figure 2: Multi-task Wasserstein Auto-encoder (MWA) independent reconstruction of $(X_1, X_2)$ minimizing \ref{['eq:multi-task-kendal']}, with $c=\mathcal{L}_{\textit{WAE}}$ and clustering of the resulting latent space, 5028 parameters.
• Figure 3: Wasserstein Auto-encoder (WA) reconstruction of $X = (X_1, X_2)$ minimizing \ref{['eq:tolstikhin2017wasserstein']} and clustering of the resulting latent space, 4932 parameters.
• Figure 4: Subset of GPS points presenting at least one BLE device reading; color map based on $e^{speed}$ shows that buses and other modes in the area have the same speed distribution--i.e., walk and bike--few trajectories recorded from car are the only exception.
• Figure 5: GPS points from smartphones, color map based on spatial density shows bus stops and bus deposit.