A Comprehensive Framework for UAV-Based Autonomous Target Finding in Environments with Limited Global Navigation Satellite Systems and Reduced Visibility

This paper presents a comprehensive framework for Unmanned Aerial Vehicles (UAVs) autonomous navigation in cluttered environments with no Global Navigation Satellite System (GNSS) and low visibility. Partially Observable Markov Decision Process (POMDP) is used to model decision making under uncertainty. The POMDP formulation was designed to allow a UAV to achieve a Search and Rescue (SAR) mission in an environment comprised of a restricted flying area, obstacles, no GNSS, and visual obscurant in the form of smoke. The mission objective is to explore the environment while avoiding obstacles to detect a human being using a thermal camera. A more realistic observation of the target's detection was modelled within the POMDP. This includes enhancing the state and observation vectors to include the characteristics of a bounding box generated by a deep-learning classifier. The framework also integrates a 2D LIDAR/Inertial odometry using the Hector SLAM package for pose estimation. It is tested in the simulation using Gazebo, Robotic Operating System (ROS), and the PX4 Firmware. Experiments conducted in the simulated SAR scenario tested the system under varying levels of pose estimation uncertainty, with an unknown target position. The experiments with the new observation function and low uncertainty pose estimation were a success. However, the framework was limited with higher uncertainty from the LIDAR/Inertial odometry, demonstrating the importance of a reliable pose estimation.