| Object detection is an important research direction in computer vision and has been widely applied in various fields such as unmanned driving,security monitoring,and disaster warning.Object detection algorithms based on deep learning networks have significant advantages over traditional machine learning methods.However,due to the diversity of objects,complexity of scenes,and high computational requirements,the performance of object detection algorithms still needs to be further improved.This paper focuses on the research of object detection in the context of unmanned aerial vehicle(UAV)aerial photography and fire scenes.Object detection in both scenarios poses certain challenges.Specifically,the challenges for UAV aerial photography include complex background information,difficult object positioning,high interference,small object pixels,and difficult feature extraction.Challenges for fire scene object detection include blurry edges and significant shape variations of flame and smoke objects,and the scarcity of publicly available datasets.This paper proposes two small object detection algorithms for UAV aerial photography and one flame and smoke object detection algorithm based on the YOLO(You Look Only Once)model as the underlying architecture,combined with a cross stage partial network,and presents a self-made flame and smoke object dataset.The specific research content of this paper is as follows:(1)We propose a small object detection algorithm,YOLO-CS,based on the Cross Stage Partial Network and Spatial Pyramid Pooling.The aim of YOLO-CS is to improve the overall accuracy of small object detection from the perspective of unmanned aerial vehicle(UAV)aerial photography.Firstly,YOLO-CS adds a CSPSPP module to the YOLOv3 backbone network,which uses a cross-stage local network and spatial pyramid pooling to increase the receptive field and obtain complex background features.Secondly,YOLO-CS introduces the idea of cross-stage local network into ordinary convolution,enhancing the network’s feature fusion and extraction abilities.Finally,YOLO-CS uses the CIoU(Complete Intersection over Union)loss function for the position function,which makes up for the deficiency of the Mean Square Error(MSE)loss function in evaluating position loss.We compare YOLO-CS with other high-performance algorithms on the Unmanned Aerial Vehicle Detection and Tracking(UAVDT)dataset.The results show that YOLO-CS achieves an average precision(m AP)of 27.05% for vehicle detection,which is 3.85% and 3.65%higher than YOLOv3 and FRCNN+FPN,respectively.Therefore,compared with the basic architecture of YOLOv3,YOLO-CS can improve the small object detection performance in simple scenarios from the perspective of UAV aerial photography by introducing a cross-stage local network.(2)A multi-scale small object detection algorithm called YOLO-CM is proposed for unmanned aerial vehicle(UAV)small object detection in complex scenarios,where objects are smaller,the environment is more complex,there are more object categories,and the dataset distribution is imbalanced.Firstly,YOLO-CM introduces two modules,namely Cross Stage Partial-Atrous Spatial Pyramid Pooling(CSP-ASPP)and Cross Stage Partial2-X(CSP2-X),into YOLOv4.The advantage of these modules is that they can obtain a larger receptive field and more edge information,thereby improving the feature representation ability of small objects.Secondly,YOLO-CM uses a k-means clustering method based on IOU to obtain matched prior boxes,instead of the statistical MS COCO dataset method used in YOLOv4,to overcome dataset mismatch.Finally,YOLO-CM adds a fourth-scale detector to the neck of YOLOv4,which can extract low-dimensional high-frequency information and is more conducive to small object detection.Performance tests of the YOLO-CM algorithm are conducted on UAVDT,Vis Drone2019(Vision Meets Drone 2019),and a self-made night-time object detection dataset called UAV-LED-G.The results show that YOLO-CM achieves an m AP of67.00% for the most representative small cars in the UAVDT dataset,which is 2.30%higher than YOLOv4.On the Vis Drone2019 dataset,YOLO-CM achieves an m AP of39.09%,which is 2.22% and 3.16% higher than YOLOv4 and DA-Retina Net,respectively.On the UAV-LED-G dataset,YOLO-CM achieves an m AP of 89.46%,which is 3.05% and 2.09% higher than YOLOv3 and YOLOv4,respectively.Therefore,the YOLO-CM algorithm can meet the requirements of small object detection in complex UAV scenarios.(3)A flame and smoke detection algorithm,YOLOX-CSP,based on a cross-stage partial network,is proposed along with a multi-scene flame and smoke dataset,FireSmokeC.Firstly,to address issues such as the variability of flame and smoke shapes and fuzzy boundaries,the algorithm combines the CSP＿ASPP and multi-scale detectors on top of the YOLOX-L architecture.Secondly,to address the scarcity of datasets for fire scenarios,a multi-scene flame and smoke dataset,FireSmokeC,was created and annotated.The YOLOX-CSP algorithm was then tested on the FireSmokeC dataset,achieving a detection m AP of 88.08%.This is a 2.14%,2.48%,and 9.78%improvement over YOLOX-L,YOLOv4,and YOLOv3,respectively,with a noteworthy detection m AP of 93.97% for flames,which is 6.26% higher than YOLOXL.These results demonstrate the high application value of the proposed algorithm for fire detection scenarios.In summary,this paper addresses two types of object detection problems: small object detection from an aerial view using drones,and object detection in fire scenarios.Three algorithms based on cross-stage partial networks and YOLO networks are proposed: YOLO-CS for small object detection from a drone’s perspective,YOLO-CM for multi-scale small object detection,and YOLOX-CSP for fire and smoke detection.Additionally,a fire and smoke object dataset is created.Extensive experiments are conducted on two publicly available drone datasets,UAVDT and Vis Drone2019,as well as two self-made datasets,UAV-LED-G and FireSmokeC,demonstrating the effectiveness and performance improvement of the proposed object detection algorithms.Therefore,these three algorithms have significant practical value in the fields of drone object detection and fire detection. |
