Object detection by deep learning approach using videos taken from unmanned aerial vehicle

Today, Unmanned Aerial Vehicles (UAV) have a wide range of uses such as border security, coast guard, defense, attack, reconnaissance, search and rescue, aerial photography, logistics, agricultural spraying, fire extinguishing. However the autonomous decision and action mechanism is a very important detail that should be emphasized, in order to avoid any intolerable delay that may occur during communication with the command center due to distance or other reasons. It is especially critical in situations such as border security, coast guard, defense, attack, fire extinguishing. The ability of UAVs to perform some tasks autonomously is can be possible by integrating the Computer Vision field with UAVs. Aerial object detection applications, which are one of the computer vision field applications, contain several errors such as not being able to detect objects of different sizes, slow detection, wrong estimation, depending on the concepts of distance and proximity. These errors are minimized by the YOLO algorithm, which learns objects of different sizes with using "Anchor Boxes" and delete bounding boxes other than the bounding box that with the maximum confidence score with using the "Non-Maximum Suppression" technique. The scope of the study is the detection of vehicles and pedestrians in the image taken from the UAV using the YOLOv3 neural network, one of the Deep Learning (DL) methods. And in this study, the video taken from the UAV was given to the neural network and the vehicles and pedestrians in the video were detected. DJI Mavic 2 Zoom Drone was used as the UAV. As dataset, 500 images taken from VIRAT [50] video datasets were used for training, and 377 images taken from the Mavic 2 Zoom videos were used for retraining the average loss value was 2,345 and the mAP value was 79%. The network was retrained with 377 images, the loss value approached 1, however, the mAP value decreased to 70.9% compared to the previous value. The training and testing process was carried out on the Google Colab Tesla T4 GPU machine. The performance of the model was also evaluated with the loss graph and mAP graph.