Anomaly Detection and Cognizant Path Planning for Surveillance Operations using Aerial Robots

Abstract

In this paper we address the problem of unsupervised anomaly detection and cognizant path planning for surveillance operations using aerial robots. Through one-class classification exploiting deep learned features on image data and a Bayesian technique to fuse, encode and update anomaly information on a real-time reconstructed occupancy map, the robot becomes capable of detecting and localizing anomalies in its environment. Provided this information, path planning for autonomous exploration of unknown areas and simultaneous maximization of the entropy of sensor observations over abnormal regions is developed. The method is verified experimentally through field deployments above a desert-like environment and in a parking lot. Furthermore, analysis results on the suitability of different deep learning-based and hand-engineered features for anomaly detection tasks are presented.