Research On Low-altitude UAV Target Detection Based On Improved SSD Detection Network

With the rapid development of artificial intelligence technology represented by deep learning,empowering various industries and broadening industry boundaries has become the trend of artificial intelligence technology development.Among them,the combination of the security industry and deep learning technology,especially computer vision,has received more and more attention.At present,due to the widespread popularity of small unmanned aerial vehicles,the threshold for the use of unmanned aerial vehicles is continuously lowered,which poses new challenges to urban security.Therefore,the detection of low-altitude UAV targets based on optical images has become a practical problem worthy of study.At present,there is no widely used standard UAV target data set in this research direction.At the same time,for the detection of UAV targets,traditional target detection algorithms are often used,and the characteristics of UAVs are not fully utilized.The accuracy needs to be improved.Besides,due to the change of low-altitude scenes,the complicated background,and the different flying attitudes of UAVs,these factors increase the difficulty of detection,but it is also more suitable for the use of deep learning-based detection technology.Therefore,this paper takes the UAV in the urban low-altitude scene as the detection target and carries out the research on the low-altitude UAV detection technology based on deep learning.Through experimental comparison,the SSD detection network is selected as the baseline.In view of the poor detection effect of this model on small targets,this paper designs an pyramid feature extraction network.At the same time,the actual receptive field technology is used to redesign the size and number of the prior boxes.By studying the imbalance of positive and negative samples in low-altitude scenes,an algorithm based on Focal Loss to gradually change the adjustment factor of difficult and easy samples is proposed,and a video detection test is performed.The work of this article mainly includes:(1)Establish a low altitude UAV target data set.Because of the currently open target detection research work,there is no universal and mature low-altitude UAV target image data set,and it is more difficult to study the deep learning-based UAV target detection method.Therefore,through shooting on the spot,drawing on existing data sets and technical simulations,this paper established a data set consisting of 20,000 positive samples of drones and negative samples of birds,kites,etc.,combined with a variety of actual scenarios,and divide the target scale.(2)Analyze the mainstream target detection framework and compare experiments.By comparing the single-stage and two-stage detection models,starting from the actual application scenarios,the SSD model that performs better in detection accuracy and detection speed is selected as the improved basic model.Its detection AP for UAVs is 80.25%,the detection speed is 0.02s/frame,which is used as the baseline of this article and focuses on improving the detection effect of small targets.(3)Through the research of low-level and high-level feature maps,in view of the problem of high-level features' weak ability to express small target features,the VGG16 low-level feature map Conv3?3 is introduced.Up-sampling the high-level feature map and adding it to the output feature map after 1×1 convolution with the previous feature map to construct a feature pyramid network,which enhances the network's feature extraction ability for small UAV.Then,by studying the relationship between the theoretical receptive field,the effective receptive field and the prior boxes in the convolution feature map,the prior boxes of different sizes and scales was redesigned,which greatly improved the detection effect for UAV,especially the small UAV.The detection accuracy of AP increased by 9.12%.Finally,the experiment analyzed the influence of each convolutional layer on the detection effect of different sizes of UAVs,and improved the customization ability of the detection network for different scenarios.(4)Aiming at the significant sample imbalance problem in UAV detection tasks in low-altitude scenarios,through studying the current advanced algorithms,this paper proposes an algorithm based on Focal Loss to gradually change the adjustment factor of difficult and easy samples,which makes difficult to distinguish the samples.The weight will not be reduced due to the increase in confidence in the training process.In comparison experiments with OHEM,GHM,and other algorithms,without reducing the detection speed,the algorithm has the highest detection AP for small target UAVs,reaching 89.98%.Experiments show that the low-altitude UAV detection algorithm based on the improved SSD model designed in this paper achieves effective detection of UAVs of various scales,and the detection AP reaches 92.84%,which is 15.69% higher than the baseline of this paper,and the detection speed is 0.031s/frame.Finally,design the video detection module and develop GUI software to test the video UAV detection in the real scene.In the hardware scenario of this paper,it meets the real-time detection requirements.