A Neuro-Symbolic AI System for Visual Question Answering in Pedestrian Video Sequences

With the rapid increase in the amount of video data, efficient object recognition is mandatory for a system capable of automatically performing question and answering. In particular, real-world video environments with numerous types of objects and complex relationships require extensive knowledge representation and inference algorithms with the properties and relations of objects. In this paper, we propose a hybrid neuro-symbolic AI system that handles scene-graph of real-world video data. The method combines neural networks that generate scene graphs in consideration of the relationship between objects on real roads and symbol-based inference algorithms for responding to questions. We define object properties, relationships, and question coverage to cover the real-world objects in pedestrian video and traverse a scene-graph to perform complex visual question-answering. We have demonstrated the superiority of the proposed method by confirming that it answered with 99.71% accuracy to 5-types of questions in a pedestrian video environment.