| Popis: |
The advent of IoT devices in combination with Human Activity Recognition (HAR) technologies can contribute to battle with sedentariness by continuously monitoring the users' daily activities. With this information, autonomous systems could detect users' physical weaknesses and plan personalized training routines to improve them. This work investigates the multimodal fusion of smartwatch sensor data for HAR. Specifically, we exploit pedometer, heart rate, and accelerometer information to train unimodal and multimodal models for the task at hand. The models are trained end-to-end, and we compare the performance of dedicated Recurrent Neural Network-based (RNN) and Convolutional Neural Network-based (CNN) architectures to extract deep learnt representations from the input modalities. To fuse the embedded representations when training the multimodal models, we investigate a concatenation-based and an outer product-based approach. This work explores the harAGE dataset, a new dataset for HAR collected using a Garmin Vivoactive 3 device with more than 17 h of data. Our best models obtain an Unweighted Average Recall (UAR) of 95.6, 69.5, and 60.8% when tackling the task as a 2-class, 7-class, and 10-class classification problem, respectively. These performances are obtained using multimodal models that fuse the embedded representations extracted with dedicated CNN-based architectures from the pedometer, heart rate, and accelerometer modalities. The concatenation-based fusion scores the highest UAR in the 2-class classification problem, while the outer product-based fusion obtains the best performances in the 7-class and the 10-class classification problems. |
