Activity Recognition with Wearable Accelerometers using Deep Convolutional Neural Network and the Effect of Sensor Placement

Human activity recognition (HAR) has become ubiquitous in modern daily life, and thus requires robust classification algorithms. Accelerometers are the most commonly used sensor for HAR, but often provide an incomplete picture of activities due to their locality. Given a sensor network of accelerometers, we believe there are two main issues that need to be addressed: (i) developing a robust, end-to-end classification framework, and (ii) identifying the optimum number and placement of sensors. To address these issues, a convolutional neural network (CNN) is implemented, tuned, and tested for activity classification. Our evaluation shows that the proposed system outperforms a number of other classifiers with a perfect classification accuracy (100%). Next, we utilize the developed pipeline to investigate the impact of different combinations of sensors and analyze HAR accuracy with respect to location and number of sensors. Our results show that at least two accelerometers are needed to achieve perfect classification for daily activities, while an accelerometer placed on the ankle is most informative for near-perfect performance.