Novel Image Dataset and Proposal of Framework for Visual Semantic Analysis Applied on Object Dangerousness Prediction

The recent technological advances, combined with the shifting and increasing necessities of the global population in the search for better commodities and ease of life, have propelled research on socially assistive robotics and artificial intelligence. Inside the gap towards complete solutions for intelligent assistive robotics is the need for more complex, autonomous, semantical knowledge-oriented approaches capable of understanding tasks. In this work, we present a novel image dataset for indoor object dangerousness targets in the context of children’s safety called the HOD Dataset. Furthermore, we introduce the concept of a Visual Semantic Analysis framework, an ongoing project for enabling visual machine-learning techniques to boost capacity by merging multiple contextual variables into a final main objective. We use the framework here to set a baseline for the HOD Dataset. The baseline is defined using a no-visual semantic analysis model, obtaining mean average precision (mAP) of 0.785 and mean average recall (mAR) of 0.838 on IoU ranging from 0.50 to 0.95.