ABSTRACT
In recent years, wearable sensors have increasingly been employed in smart health applications. Wearable sensors can not only be used to collect users’ valuable health-related data, they can be used in conjunction with other infrastructure-bound sensors, such as Microsoft Kinect sensors, to facilitate privacy-aware fine-grained activity tracking. This fusion of multi-modal data promises a new type of smart health application that coaches a user to live a healthier lifestyle by monitoring the user in real time and reminding him or her when he or she engages in an unhealthy activity. In this article, we investigate how to achieve fine-grained activity recognition in the context of such an application. In our scheme, identification accuracy is improved by incorporating a nonlinear local similarity measure, namely kernel risk-sensitive loss (KRSL), into a novel multilayer neural network (NN) learning algorithm, a stacked extreme learning machine (S-ELM). Furthermore, to achieve good generalization performance with minimal human intervention, the popular optimization algorithm Jaya is also used to adjust key parameters in our proposed approach. Experiments were conducted to verify the effectiveness of the proposed scheme.
