Visual object tracking by using deep neural networks

In this thesis, new methods have been proposed to track arbitrary objects. To this end, first, a novel lightweight two-stream deep learning-based tracker has been developed to obtain state-of-the-art results on different single object tracking benchmarks. In the proposed deep learning-based method, both spatial and temporal features are used and a ranking loss is employed. Using ranking loss term in the optimization function enforces the neural networks to learn to give higher scores to the candidate regions that better frame the target than the regions that frame the target object with less accuracy. The classifier scores received from spatial and temporal networks are fused. The experiments were conducted on eight different benchmarks. The proposed tracker achieves state-of-the-art results on most of the tested challenging tracking benchmarks. In addition to a deep learning-based tracker, a hybrid correlation filter-based tracker has been developed to detect and track UAVs. To detect the target UAV at the beginning of the tracking and in the case where the tracked UAV has been lost, the deep learning based YOLOv3 detector has been used. A kernelized correlation filter has been used to track detected target UAVs in real-time. Combining the detector and tracker provides high accuracy and real-time performance even on onboard computers. A new dataset called as UAV Tracking has been created to train the YOLOv3 detector and test the proposed method. The proposed method achieves state-of-the-art performances on created UAV Tracking dataset. Finally, the proposed method has been generalized for general tracking.