A Lightweight Adaptive Attention Based Transportation Mode Detection Model Using Sensors Integrated in Smartphones

In transportation planning, user transportation mode is a significant part of traffic survey and an important basis for traffic design. However, how to obtain user transportation mode faces great challenges due to the lack of expert knowledge, high energy consumption, and limited usage scenarios. To tackle these challenges, in this paper, we propose a lightweight adaptive attention-based transportation mode detection model using sensors integrated in smartphones. We adopt Multi-Head Self-Attention (MHSA) mechanism and Fully Connected Residual Network (FR) in feature learning to improve model accuracy. Meanwhile, MHSA is highly parallelizable and has a compact parameter set, making it more suitable for mobile edge devices with limited device memory and computation capability. Adding residual connections to the model can improve its information transfer efficiency and reduce its training time. Extensive experiments on SHL datasets demonstrate that our proposed model outperforms five state-of-the-art baselines and achieves 4.20% higher accuracy than the best baseline on SHL dataset.