Abstract:

Today, Unmanned Aircraft has a wide range of uses such as border security, coastal security, defence, attack mainly search rescue, cranberry medication, fire-extinction. However, if IHA can perform some tasks autonomously, it is with the integration of the computer vision system. One of its practices in this field, air-to-air object detection practices may include a variety of errors, such as failure to detect objects of different sizes depending on the concepts of distance, proximity, slow detection, misprediction. Those errors can be minimized by deep learning practices. In this study, the YOLOv3 DÖ network was trained with the images taken from the VRAT[1] video set and then re-trained with the images taken from the DJI Mavic 2 Zoom IHA camera, which provided the identification of the vehicles and trails in the video. The training and test process was carried out on the Google Colab Tesla T4 GPU machine. The model’s performance for the first and second training was Loss, mAP 2.345, 79% and 1.171, 70.09%.

Keywords: