Research On Drones Detection Algorithm Based On Deep Learning

Drones are a great invention in the 21 st century.With the development of technology,drones,especially civilian drones,have developed rapidly and are widely used in various fields.The application of drones has greatly reduced the cost of working at heights,making many tasks previously inconvenient for humans to accomplish easily.While drones bring convenience to people's lives,they also bring many potential threats to the country and society.Drones are small in size and fast in flight,and have unique advantages in the field of surveillance and video shooting,but this has brought great problems to the supervision of drones.The traditional drone detection algorithm has high cost and low accuracy,which is not suitable for large-scale use.In this master thesis,an algorithm based on deep learning is used for drone detection.Drone detection based on deep learning can overcome the shortcomings of traditional algorithms and improve the detection capabilities of drones.At present,in China,deep learning-based drone detection technology is still in the preliminary stage.Therefore,this master thesis analyzes the object detection algorithm and instance segmentation algorithm based on deep learning and develops the following work for the characteristics of small drone and fast flight speed:(1)For drone data sets,deep learning is a data-based solution,but commonly used data sets do not include drones.This master thesis considers that there is a huge difference in the shape of military drones and civilian drones,and selects civilian multirotor drones to make data sets.Considering that the edge of the rotor is difficult to shoot clearly when the multi-rotor drones is non-static,this master thesis only labels the drones rotor with clear edges.This master thesis uses Labelme software to label the drone,make a data set that contains masks for instance segmentation,and convert the data set to COCO format to complete instance segmentation training and convert to VOC data set format to convenient to complete the task of object detection.(2)Aiming at the objects with small volume and high speed,such as drones,this master thesis proposes an improved YOLOv3(You Only Look Once version 3)algorithm based on QCA(Quotient of Circumcircle Area)loss function,and is trained to adapt to the detection of drone.The flying speed of the drones is fast,and the time spent in the video screen is short.The calculation cost of the commonly used object detection algorithm is large,and it is difficult to complete this task.Based on the YOLOv3 algorithm,this master thesis analyzes the problem of inaccurate positioning of the prediction boxes caused by the YOLOv3 object detection algorithm due to the inappropriate loss function during the convolutional neural network training process.A QCA loss function is proposed to improve the accuracy of the detection of the prediction boxes and solve the problem of poor detection effect of the YOLOv3 algorithm in the detection process of smaller objects.(3)This master thesis attempts to use the instance segmentation algorithm to complete the drone semantic segmentation while completing the drone detection.However,the general instance segmentation algorithm may have problems such as missing detection for small objects such as drones.In this master thesis,the YOLACT algorithm is selected as the basis,and the backbone network HRNet(High-Resolution Representation Net)with high-resolution representation for attitude detection is integrated,and the HR-YOLACT algorithm is proposed.The HR-YOLACT algorithm considers the computational cost of the network,and uses HRNet-W40 as the backbone of the network to extract image features.So that the instance segmentation algorithm can obtain more detailed features in the feature extraction stage.In this master thesis,the input layer of the feature pyramid is reduced to obtain better In-depth information,and the algorithm detection head is modified to improve the algorithm's object detection ability,so as to improve the detection ability of the drone target.HR-YOLACT algorithm improves the detection ability of the algorithm through the above operations without significantly increasing the calculation cost.