Multi-Domain Indoor Dataset for Visual Place Recognition and Anomaly Detection by Mobile Robots

Abstract

Visual location recognition encompasses place recognition (PR) and anomaly detection (AD). These are crucial tasks for autonomous robots to accurately determine the location and the occupied place. To accelerate research in this area, we introduce a multi-domain dataset for indoor visual place recognition and anomaly detection by mobile robots. The dataset includes 89,550 RGB images captured in nine rooms. The data collection process involved both manual recordings and recordings captured by mobile robots. The images depict a wide range of scenarios, including variations in lighting, robot vision, and human activity. Additionally, we provide an analysis of other available datasets referenced in the literature. This article presents a freely available dataset for research on place recognition and presents an example application in the field of anomaly detection. The baseline methods were thoroughly tested and achieved an 80.18% accuracy in anomaly detection for single images and 80.63%-84.18% for image sequences. The article includes a comprehensive presentation of the characteristics of individual image sequences and the most significant conclusions drawn from the research.