GLMLP-TRANS: A transportation mode detection model using lightweight sensors integrated in smartphones

Abstract

Transportation mode detection (TMD), as an essential part of Intelligent Transportation Systems, aims at analyzing human current transportation activities, and can be widely adopted in road planning and traffic prediction. Although several relevant works have been conducted on TMD, there are still some great challenges due to complex factors including: (i) availability that several sensors are limited in certain scenarios; (ii) lightweight that both data collection sensors and the architecture of most current neural network are heavy and sophisticated; and (iii) expert knowledge that numerous researches for TMD are based on traditional machine learning. To address these challenges, we propose a deep learning-based transportation mode detection network using smartphone-based sensors called GLMLP-TRANS, which is inspired by Self-attention and MLP-Mixer. Our proposed GLMLP-TRANS network can capture both global and local temporal features and adaptively combine them to learn comprehensive information of each transportation mode. Residual technique is introduced to accelerate the progress of learning and enhance the accuracy of transportation mode detection. Extensive experimental results on SHL dataset demonstrate that our proposed GLMLP-TRANS network outperforms other six baselines and improve 12.8% than the best baseline on SHL dataset.