Application of Deep Learning Based Object Detection on Unmanned Aerial Vehicle

Abstract

Object recognition can be seen in every field thanks to today's high computation power. It has become possible to extract information from the images with high accuracy in different areas such as human, animal, vehicle, traffic signs and MR images. In this study, plastic water bottle detection was applied by using Inception v2, MobileNet v2 and ResNet50 architectures to be used in the Scientific and Technological Research Council of Turkey (TUBITAIC) International Unmanned Aerial Vehicle Competition. Cloud based training applied to locate very small objects from up to 5 meters high. In the present study, the Single Shot Detector (SSD) ResNet50 model yielded the best results for mAP@1.75 and final loss. It was also determined that SSD Inception v2 model gave better results than SSD MobileNet v2 model. When the processing times were considered, it was determined that models could not work at Raspberry Pi 3 Model B+ equipment at sufficient speed. However, these results have shown that the Jetson Nano card can be operated correctly on a drone. In the next stage of this study, the aim will be determining environmental pollution level by detecting objects such as plastic bags, bottles and fruits on the streets.