Real world performance of an RNN model for sustainable mobility metrics

This is the authors' abstract. We don't add key points for this paper.

This study presents the development, deployment, and testing of a Recurrent Neural Network (RNN) model for real-time transportation mode classification using smartphone sensor data. The system integrates a mobile app and cloud-based backend, enabling continuous data collection and model inference. Experimental results show that the model achieves high categorical accuracy (91.8%) and F1-score (78.0%) on the test set, with excellent precision and recall for common modes such as car , foot , and still . The model inference takes ∼ 130 ms per block of points. However, classification of the metro mode remains challenging due to limited GPS availability in underground environments. The analysis highlights that GPS-based data triggering is suboptimal for such contexts, suggesting the adoption of Human Activity Recognition (HAR) state transitions as a more effective alternative. Future improvements will focus on optimizing data collection strategies and extending the dataset to multiple urban areas to improve generalizability. These findings contribute to the advancement of mobile transport mode recognition for transportation behavior analysis and policy-making. • The model achieved 91.83% categorical accuracy in predicting transportation modes. • High precision was observed for car (93.4%), foot (97.7%), and still (85.1%) classes. • Performance was lower for the metro class due to the architectural limitations. • A HAR-based triggering strategy would improve both efficiency and classification. • The model was deployed in the cloud backend, paired with a mobile app frontend.