Deep Learning-Based Robotic Perception for Adaptive Facility Disinfection

Hospitals, schools, airports, and other environments built for mass gatherings can become hot spots for microbial pathogen colonization, transmission, and exposure, greatly accelerating the spread of infectious diseases across communities, cities, nations, and the world. Outbreaks of infectious diseases impose huge burdens on our society. Mitigating the spread of infectious pathogens within mass-gathering facilities requires routine cleaning and disinfection, which are primarily performed by cleaning staff under current practice. However, manual disinfection is limited in terms of both effectiveness and efficiency, as it is labor-intensive, time-consuming, and health-undermining. While existing studies have developed a variety of robotic systems for disinfecting contaminated surfaces, those systems are not adequate for intelligent, precise, and environmentally adaptive disinfection. They are also difficult to deploy in mass-gathering infrastructure facilities, given the high volume of occupants. Therefore, there is a critical need to develop an adaptive robot system capable of complete and efficient indoor disinfection. The overarching goal of this research is to develop an artificial intelligence (AI)-enabled robotic system that adapts to ambient environments and social contexts for precise and efficient disinfection. This would maintain environmental hygiene and health, reduce unnecessary labor costs for cleaning, and mitigate opportunity costs incurred from infections. To these ends, this dissertation first develops a multi-classifier decision fusion method, which integrates scene graph and visual information, in order to recognize patterns in human activity in infrastructure facilities. Next, a deep-learning-based method is proposed for detecting and classifying indoor objects, and a new mechanism is developed to map detected objects in 3D maps. A novel framework is then developed to detect and segment object affordance and to project them into a 3D semantic map for precise disinfection. Subsequently, a novel deep-learning network, which integrates multi-scale features and multi-level features, and an encoder network are developed to recognize the materials of surfaces requiring disinfection. Finally, a novel computational method is developed to link the recognition of object surface information to robot disinfection actions with optimal disinfection parameters