Research And Implementation Of Drone Detection Algorithm Based On Deep Learning

At the same time as the rapid development of the drone industry,it also brings challenges to people’s security and privacy.It is urgent to supervise drones,and drone detection is the key to implementing supervision.Detecting drones based on traditional methods has the characteristics of high cost and complicated implementation,which are not suitable for large-scale promotion.Therefore,this article focuses on the efficient detection of drones based on deep learning.Aiming at the problem that the drone detection data set is missing,a drone detection data set is constructed.Firstly,get high-quality drone images from drone motion videos.Secondly,each image is labeled with its position and category,where the position is represented by the pixel coordinates of the upper left and lower right of the rectangular frame surrounding the drone.Finally,it is constructed with reference to the standard target detection data set.Aiming at the performance gap between the current general target detection and drone detection,the implementation steps and development process of current target detection are analyzed.Firstly,the traditional target detection method is explained,and its existing problems are obtained.Secondly,systematically analyze the implementation process of the target detection algorithm based on deep learning,and find the basic method for realizing drone detection.Finally,researches were focus on one-step detection algorithms and two-step detection algorithms that are widely used in industry,and provide theoretical basis for the improvement of detection speed and accuracy of drones.Aiming at the problem of improving the accuracy of drone detection,an improved drone detection algorithm based on Faster RCNN is proposed to achieve higher detection accuracy,which is combined the advantage of the accuracy advantage of the two-step detection algorithm with the feature pyramid’s ability to extract small target drone features.Firstly,the design pyramid of the feature pyramid and the basic feature extractor is used as the input of the RPN network,so that the drone candidate area obtained has richer semantic information of the drone.Secondly,applying ROI Align to candidate target alignment solves the problem of target position shift caused by Ro I pooling.Finally,both the self-built drone data set and the PASCAL VOC data set achieved high accuracy,which proved the effectiveness of the algorithm improvement.Aiming at the problem that the detection accuracy and detection speed of the drone do not match,an improved drone detection algorithm based on YOLOv3 is proposed to ensure real-time detection while achieving higher detection accuracy,which is combined the advantage of the detection speed of the one-stage target detection algorithm with the loss function that is more suitable for the detection of the drone.Firstly,for the problem of inaccurate description of the drone position,the Generalized intersection over union is used to measure the gap between the predicted target and the actual target,which makes the positioning of the drone more accurate.Secondly,a serious imbalance of positive and negative samples appeared during the training process,which is caused by the small target drone occupying a small position in the image.So,focus loss was used to suppress the negative optimization problem caused by negative samples.Finally,experiments show that the algorithm proposed in this paper has obvious advantages in detection accuracy and detection speed,and can achieve realtime detection of drones,which has a strong application value.