Provide a Method Based on Deep Learning Networks to Detect Drones from Depth Data

Today, drones have received much attention due to their many benefits and wide range of applications such as mapping, agriculture, and crisis management, and have replaced many of the existing traditional methods. However, the expansion of drones and their unauthorized entry into important infrastructure such as government buildings can pose potential threats to public safety. Therefore, small drone detection, localization, and tracking systems are critical. In recent years, neural network-based diagnostic methods and deep learning approaches have shown considerable ability in the field of drone detection. Therefore, in this research, a detection method based on deep convolutional learning networks has been used to detect drones. On the other hand, the use of visible images faces problems such as the presence of hidden areas, crowded backgrounds, and the impossibility of separating the background and light problems inside the image. Also, thermal image-based systems, despite having night vision power, have less spatial resolution than the visible image, which makes the drone detection process difficult. Recently, the use of depth images, which do not have the challenges associated with visible images and show the distance and proximity of the object to the camera, has become very popular. In this research, drone detection has been performed using a collection of simulation and real depth images and using YOLO (You Only Look Once) deep learning network. The real depth images in this study were obtained through a Semi Global Matching algorithm (SGM), and finally, the validation of the trained model was examined with a variety of real-time and simulation depth images with three types of drones inside and outside the environment. Finally, the results of drone detection with the desired deep learning network in the simulation depth images reached an average precision of 84%, an average time of 125 ms, and in real depth images an average precision of 74%, an average time of 133 ms.