Towards Real Time Interpretable Object Detection for UAV Platform by Saliency Maps
Maxwell Hogan, Nabil Aouf · 2021 IEEE International Conference on Robotics and Biomimetics (ROBIO) · 2021
On-board object detection is an important requirement for Unmanned Aerial Vehicles (UAVs) when carrying out a variety of tasks such as obstacle avoidance, search and rescue, and automatic target recognition. One of the main difficulties of conducting object detection for a UAV is that, because the objects of interest are observed from altitude, this causes them to appear very small in images acquired from an onboard camera. In this work, we attempt to provide a solution to this difficulty, while also seeking to meet two other criteria that are important to the deployment of Artificial Intelligence with UAVs: firstly, the capability to operate in real-time; and secondly, the ability of the user to be able to trust the predictions it makes. To meet the challenge of small object detection we present the use of an image Tile Loader to enhance the capability of Deep Neural Network (DNN) style detectors, while minimising the processing time costs. Furthermore, we also introduce the practice of using Grad-CAM to provide better insights into a detection style architecture as a means to enhance trustworthiness.