Indoor monitoring of people at their homes has become a popular application in Smart Health. With the advances in Machine Learning and hardware for embedded devices, new distributed approaches for Cyber-Physical Systems (CPSs) are enabled. Also, changing environments and need for cost reduction motivate novel reconfigurable CPS architectures. In this work, we propose an indoor monitoring reconfigurable CPS that uses embedded local nodes (Nvidia Jetson TX2). We embed Deep Learning architectures to address Human Action Recognition. Local processing at these nodes let us tackle some common issues: reduction of data bandwidth usage and preservation of privacy (no raw images are transmitted). Also real-time processing is facilitated since optimized nodes compute only its local video feed. Regarding the reconfiguration, a remote platform monitors CPS qualities and a Quality and Resource Management (QRM) tool sends commands to the CPS core to trigger its reconfiguration. Our proposal is an energy-aware system that triggers reconfiguration based on energy consumption for battery-powered nodes. Reconfiguration reduces up to 22% the local nodes energy consumption extending the device operating time, preserving similar accuracy with respect to the alternative with no reconfiguration.