Research On Small Target Intelligent Detection Technologies For UAV Inspection

With the rapid development of Internet of Things technology,industrial intelligence has become an important trend in modern industrial development.The unmanned aerial vehicle(UAV)inspection,a new emerging application technology,has broad prospects in improving industrial intelligence.UAV inspection has various sensing and decision-making methods,and visual object detection technology is considered as one of the main methods due to its efficiency and flexibility.However,existing object detection algorithms designed for natural scenes cannot achieve satisfactory results when directly applied to small object detection in UAV aerial images,due to problems such as multi-scale variation of targets and a large proportion of small targets.To address these issues,this thesis proposes two small object detection algorithms for UAV aerial images.The main work is as follows:(1)Firstly,an overview is given of UAV inspection technology and object detection technology.Then,detailed introductions are provided on image pre-processing techniques,object detection techniques based on convolutional neural networks,and object detection evaluation metrics.(2)A UAV inspection traffic object detection algorithm based on adaptive image defogging is proposed.Firstly,to address the issue of the low detection accuracy of small traffic objects by the basic YOLOv5 network in foggy aerial scenes,a dual corrective adaptive dark channel prior defogging algorithm is proposed to maximize the restoration of traffic object details,which can adaptively dehaze and correct brightness for images with varying fog densities.Then,a proposed hierarchical weighted spatial pyramid pooling structure is used to replace the original fast spatial pyramid pooling structure,which enhance the feature extraction capability for small objects.Finally,fine-tuning is performed on the preprocessing,feature fusion structure,and post-processing parts of the YOLOv5 network based on the characteristics of small traffic objects in aerial images.Experimental results show that the proposed algorithm can achieve accurate detection of traffic targets in UAV aerial images under foggy inspection scenarios.(3)A forest fire detection algorithm for UAV inspection based on binary collaborative feedback is proposed.Firstly,to address the issue of the low detection accuracy of smoke and flame objects in aerial images by the basic FCOS network,the improved network structure and training strategy of the FCOS model is proposed.Based on this,a novel binary detection network based on the improved FCOS network is further constructed to improve the detection accuracy of smoke and flame objects.Then,to address the issue of weak detection accuracy in continuous video detection using the proposed binary detection network,two collaborative feedback mechanisms are introduced to adaptively adjust the input of the next frame or the fusion weights of the network,which strengthens the continuous detection capability of the binary detection network for smoke and flame targets in videos.Experimental results show that the proposed algorithm can achieve continuous and accurate detection of smoke and flame targets in UAV aerial videos for forest fire inspection.