Anomaly detection and tracking for a patrolling robot

Abstract

This paper presents a mobile robot capable of repeating a manually trained route and detects any visual anomalies that were not present during the training run. A monocular panoramic vision sensor is used for both repeating the route and anomaly detection. Anomalies are detect by extracting the differences between images captured during autonomous runs and reference images captured during training runs. This enables detection of both mobile anomalies, such as an intruder, and stationary anomalies, eg. a suspicious suitcase. However, routes are not repeated exactly and small deviations from the reference route leads to differences in image appearance. Two stereo correspondence algorithms are used to mitigate this problem and a performance comparison using manually segmented ground truth is performed in this paper. Anomalous regions are subsequently tracked using a particle filter. Experiments in three different environments are presented, a cluttered robotics laboratory, a corridor, and an office environment. 1