Abstract:

Unmanned aerial vehicles (UAVs), called aerial robotics, have been frequently used recently to collect data and images in civilian and military areas such as disaster management, traffic density and border security. Due to the variability of the camera view angle, there are difficulties in making instant object detection on the images taken at high altitude with the UAV. This study aims to detect objects from a camera connected to the UAV with the CSI (Camera Serial Interface) module using images taken at different angles and conditions and a Convolutional Neural Network-based SSD MobileNet library. The images obtained with the camera via the UAV were processed and classified with the NVIDIA Jetson Nano embedded system computer. One of the problems encountered in real-time target detection is the classification of moving objects with low resolution due to variable weather conditions and lighting. A solution to this situation is sought by matching the close-featured data between the learned images contained in the deep learning library and the images from the camera. In this study, the images obtained by creating an algorithm in the embedded system were edited, and the outputs obtained after the feature comparison process with the deep learning library were evaluated using the object boundary and mAP (mean Average Precision) average precision percentage. As a result, 95.5% mAP final accuracy and 69.45% Map average accuracy for human detection, 83.4% mAP final accuracy and 64.5% mAP average accuracy for vehicle detection were obtained from real-time visual data received from the UAV.

Keywords: