Deep Learning Object-Recognition in a Design-to-Robotic-Production and -Operation Implementation

This paper presents a new instance in a series of discrete proof-of-concept implementations of comprehensively intelligent built-environments based on Design-to-Robotic-Production and -Operation (D2RP&O) principles developed at Delft University of Technology (TUD). With respect to D2RP, the featured implementation presents a customized design-to-production framework informed by optimization strategies based on point clouds. With respect to D2RO, said implementation builds on a previously developed highly heterogeneous, partially meshed, self-healing, and Machine Learning (ML) enabled Wireless Sensor and Actuator Network (WSAN). In this instance, a computer vision mechanism based on open-source Deep Learning (DL) / Convolutional Neural Networks (CNNs) for object-recognition is added to the inherited ecosystem. This mechanism is integrated into the system's Fall-Detection and -Intervention System in order to enable decentralized detection of three types of events and to instantiate corresponding interventions. The first type pertains to human-centered activities / accidents, where cellular- and internet-based intervention notifications are generated in response. The second pertains to object-centered events that require the physical intervention of an automated robotic agent. Finally, the third pertains to object-centered events that elicit visual / aural notification cues for human feedback. These features, in conjunction with their enabling architectures, are intended as essential components in the on-going development of highly sophisticated alternatives to existing Ambient Intelligence (AmI) solutions.