Title: Toward Building Safer Smart Homes for the People with Disabilities

Abstract: Situational awareness is a critical foundation for the protection of human life/properties and is challenging to maintain for people with disabilities (i.e., visual impairments and limited mobility). In this paper, we present a dialog enabled end-to-end assistive solution called "SafeAccess" to build a safer smart home by providing situational awareness. The key functions of SafeAccess are: - 1) monitoring homes and identifying incoming persons; 2) helping users in assessing incoming threats (e.g., burglary, robbery, gun violence); and, 3) allowing users to grant safe access to homes for friends/families. In this work, we focus on building a robust model for detecting and recognizing person, generating image descriptions, and designing a prototype for the smart door. To interact with the system, we implemented a dialog enabled smartphone app, especially for creating a personalized profile from face images or videos of friends/families. A Raspberry pi connected to the home monitoring cameras captures the video frames and performs change detection to identify frames with activities. Then, we detect human presence using Faster r-cnn and extract faces using Multi-task Cascaded Convolutional Networks (MTCNN). Subsequently, we match the detected faces using FaceNet/support vector machine (SVM) classifiers. The system notifies users with an MMS containing the name of incoming persons or as "unknown", scene image, facial description, and contextual information. The users can grant access or call emergency services using the SafeAccess app based on the received notification. Our system identifies persons with an F-score 0.97 and recognizes items to generate image description with an average F-score 0.97.