Adaptive detection tracking system for autonomous UAV maritime patrolling

Nowadays, Unmanned Aerial Vehicles (UAVs) are considered reliable systems, suitable for several autonomous applications, especially for target detection and tracking. Although significant developments were achieved in object detection systems over the last decades using the deep learning technique known as Convolutional Neural Networks (CNN), there are still research gaps in this area. In this paper, we present a new object detection-tracking algorithm that can be used on low power consuming processing boards. In particular, we analysed a specific application scenario in which a UAV patrols coastlines and autonomously classifies different kind of marine objects. Current state of the art solutions propose centralised architectures or flying systems with human in the loop, making the whole system poorly efficient and not scalable. On the contrary, applying a Deep Learning detection system that runs on commercial Graphics Processing Units (GPUs) makes UAVs potentially more efficient than humans (especially for dull tasks like coastline patrolling) and the whole system becomes easily scalable because each UAV can fly independently and the Ground Control Station does not represent a bottleneck. To deal with this task, a database consisting of more than 115000 images was created to train and test several CNN architectures. Furthermore, an adaptive detection-tracking algorithm was introduced to make the whole system faster by optimizing the balancing between detecting new objects and tracking existing targets. The proposed solution is based on the measure of the tracking confidence and the frame similarity, by means of the Structural SIMilarity (SSIM) index, computed both globally and locally. Finally, the developed algorithms were tested on a realistic scenario by means of a UAV test-bed.