Research On Aerial Object Detection Algorithm Based On Deep Learning

The importance of unmanned aerial vehicle(UAV)based object detection in various fields such as agriculture,industry,and security are increasingly evident.However,UAVbased object detection faces significant challenges due to complex backgrounds,high image resolutions,small target sizes,large scale variations,and limited computational capabilities of deployment platforms.To address the issues of scale variations,complex backgrounds,and low detection accuracy caused by small targets,this paper focuses on researching real-time object detection algorithm YOLO,with the aim of constructing a more balanced approach in terms of accuracy and speed for UAV-based object detection.The main contributions and achievements of this study are as follows:(1)Construction and augmentation of an infrared aerial pedestrian dataset.Compared to visible light datasets,the current infrared aerial pedestrian datasets suffer from limitations such as insufficient data volume and limited scenes.By equipping a UAV with an infrared camera,a new infrared aerial pedestrian dataset,named FLIR-UAV,was constructed and expanded.This dataset consists of 3302 images captured by the UAV at low flight altitudes from five typical scenes.(2)A YOLOv5-based aerial object detection algorithm,named CECA-YOLOv5,was developed by integrating ConvNeXt and ECA attention mechanisms.To improve the detection capability for small objects,a detection head corresponding to the shallow stage of YOLOv5 s was added.Furthermore,a bidirectional feature pyramid network(Bi FPN)was introduced in the neck network to enhance the feature aggregation capability,and a new C3 module called CECA,which combines ConvNeXt and ECANet designs,was proposed to further enhance the algorithm’s accuracy.The experiments were conducted using VEDAI,NWPU VHR-10,Vis Drone,and FLIR-UAV dataset.The ablation experiments on the VEDAI dataset demonstrated the comparative advantages of the CECA-YOLOv5 algorithm over mainstream detection algorithms such as YOLOv5,YOLOR,and YOLOX.For example,compared to YOLOv5 s,CECA-YOLOv5 achieved improvements of 5.5% in m AP and 6.5% in m AP50.(3)An aerial object detection algorithm,MEAPF-YOLOv5,was proposed by incorporating the ECANet into the modified spatial pyramid pooling(SPP)structure.Firstly,the CECA module was utilized to enhance the feature aggregation capability of the neck C3 module in YOLOv5.Secondly,to further improve the multi-scale feature extraction capability,the ECANet was integrated into the improved SPPF structure to create the MEAPF structure.The application value of MEAPF-YOLOv5 was validated on the VEDAI,NWPU VHR-10,Vis Drone,and FLIR-UAV datasets.For instance,when evaluated on the VEDAI dataset,MEAPF-YOLOv5 achieved comparable m AP50 performance with a parameter size and GFLOPs of only 19% and 11.9%,respectively,compared to the large-scale mainstream algorithm ECAP-YOLOv5 l.(4)Deployment and optimization of CECA-YOLOv5 and MEAPF-YOLOv5 algorithms on the Jetson TX2 platform using TensorRTwere studied,considering limited memory and power consumption of onboard embedded systems.While maintaining inference accuracy,the inference time was reduced by approximately 60%,effectively enhancing the practical application value of the aerial object detection algorithms for UAV onboard devices.