Attention Mechanism-Based Bidirectional Long Short-Term Memory for Cycling Activity Recognition Using Smartphones

Abstract

Bicycles are an ecofriendly mode of transportation, and cycling offers physical and mental well-being. However, their increased use has resulted in frequent bicycle–human accidents, car-to-bicycle collisions, related injuries and cyclist crashes. Moreover, rules for safe cycling are limited. Smart healthcare systems using smartphones and/or wearable devices, such as a cycling monitoring application that can inform fellow cyclists about the state of the user, can be developed to provide assistance during such unexpected events. In this study, a one-dimensional convolutional neural network (1DCNN)–bidirectional long short-term memory (BiLSTM) based on attention mechanism (CBiAM) model is proposed for detecting cyclists’ states using a mobile phone, thereby enhancing their safety and promoting a secure cycling experience in case of accidents or emergencies. In addition, the “cycling safe (CySa) dataset,” a new dataset containing data on the cyclists’ actions during cycling, collected from a smartphone positioned in the cyclists’ pocket is presented. The proposed CBiAM model was trained on the CySa dataset using different sliding window sizes, batch sizes (Bz), and learning rates (Lr). Experimental results confirmed the superior performance of the proposed model compared to conventional approaches, such as support vector machines and artificial neural networks, and existing advanced architectures, such as 1DCNN, long short-term memory (LSTM), and Bi-LSTM. The robustness of the model was validated using public datasets, such as UCI-human activity recognition (HAR), PAMAP2, Opportunity, MOTIONSENSE, and WISDM, where it achieved impressive F1-scores of 97.51%, 99.82%, 94.72%, 97.67%, and 87.05%, respectively.

Keywords

• Attention mechanism
• cycling
• human activity recognition
• bidirectional LSTM