Object tracking applications for intelligent control systems in unmanned aerial vehicles

Unmanned aerial vehicles are frequently used nowadays and the demand for them is increasing. As the aerial vehicles are becoming unmanned, the need for autonomous control applications increases rapidly. Unmanned aerial vehicles are used in a variety of conditions and for a variety of tasks, requiring intelligent control applications for full autonomous flight. In an intelligent control system, control systems such as various machine learning, deep learning and statistical learning methods have been developed and continue to be developed. As it is known, unmanned aerial vehicles can move very fast. When an unmanned aircraft with obstacle avoidance or object tracking is considered, the intelligent control system that provides control must make decisions very quickly and accurately, otherwise the aircraft will either hit the obstacle or lose the object by making the wrong decision. For this, the object tracking algorithm to be developed must have two essential features. These features are speed and accuracy. The speed of the algorithm depends on the computing power of the algorithm to be used and the adequacy of the hardware to be used. The accuracy of the algorithm will depend on how accurately the method used makes the classification or track the object. For this reason, in this study, the performance of deep learning and statistical learning methods used in the literature is compared and the most suitable method for object tracking for intelligent control systems in unmanned aerial vehicles is tried to find.For this study Particle Filter (PF), Kalman Filter (KF), Faster R-CNN and Single Shot Multibox Detector (SSD) are used in order to track objects during stable and moving flights on 5, 10 ,20 meter altitude. After this the results were compared and teh most suitable object tracking method is aimed to find. For the purpose of performance evaluation, frames per second (FPS), tracking accuracy, GPU and CPU usage percentages of each method during each flight were compared. When the results were considered, Faster R-CNN had the highest success rate, followed by SSD, KF and PF, respectively. When FPS ratios were compared, KF yielded the best results and PF, SSD and Faster R-CNN respectively. SSD and PF were very close in FPS rates. As a result of the comparisons, the most suitable method for tracking objects in smart control systems for unmanned aerial vehicles is observed as SSD by considering 12 FPS speed and average %90 success rate on lower altitude. However, in systems where cost, system weight and computing power should be low, it is concluded that the use of KF will be appropriate.