Implementation of Target Tracking Methods on Images Taken from Unmanned Aerial Vehicles

Traditional object detection algorithms generate proposals and implement feature extraction. Then, a classification algorithm is implemented to label object classes. This process is slow, and the accuracy may not be adequate for UAV's real-time application tasks due to their movement in the air. We specified and practically implemented an object detection and localization scheme on images taken from a UAV, and provided the UAV with an advanced vision. We used YOLOv2 model. The YOLOv2 is a suitable object detection approach based on deep learning, and it presents a network architecture with accurate results in high speed. The object detection and localization were successfully implemented for people, car, and motorcycle classes within the threshold confidence scores. We pre-trained the model on COCO dataset and tested the model with our test images. The confidence scores were higher in altitudes from 5 to 15 meters and the confidence scores varied between %45 - %79 mAP.