Scalable Passenger Detection Using Smartphone–Bus Implicit Interactions

To address these issues, we build on the concept of implicit be-in/be-out (BIBO) smartphone sensing and classification, introducing a platform that collects Bluetooth Low Energy (BLE) signals from devices on buses and GPS data from both buses and smartphones. We propose a cause–effect multitask Wasserstein autoencoder (CEMWA) architecture to train a model using GPS features and BLE signals as mutual pseudolabels. CEMWA integrates various frameworks around Wasserstein autoencoders and neural networks, providing a validated latent space representation of users’ smartphones within the transport system. This representation facilitates BIBO clustering via density-based spatial clustering of applications with noise.

Our comparative study of CEMWA’s architecture and benchmarking against best-in-class supervised methods reveals that, while Extreme Gradient Boosting and the random forest are robust to label noise, CEMWA’s design inherently handles label noise, achieving the best performance, with an 88% F₁ score in a BIBO scenario.