Small Object Detection Algorithm From The UAV Based On Attention Mechanism And YOLOv5 Network

With the rapid development of artificial intelligence algorithms and computing power,object recognition based on deep learning frameworks has become the main research direction in the field of computer vision.At the same time,drones equipped with object detection terminals have been gradually applied in many fields,such as urban security,intelligent transportation,disaster warning,emergency rescue,and cultural tourism.However,due to complex factors such as environment,lighting,height,and angle,images taken from the perspective of drones have problems such as small object size,high object density,low clarity at night,easy obstruction,imbalanced categories,and large background interference,making drone object detection tasks challenging.This article focuses on unmanned aerial vehicles(UAV)object detection and proposes the YOLO-DUAV algorithm,which combines attention mechanism and YOLOv5 architecture for detecting small objects on drones and completes performance verification.It also constructs the first drone object detection database under siliconbased golden light LED source and completes performance testing of multiple algorithms on the database.Finally,a lightweight drone object detection algorithm YOLO-UAVLight oriented towards edge computing is proposed,and the algorithm is deployed on embedded terminals.The main research contents of this article are as follows:(1)The YOLO-DUAV algorithm based on multi-attention mechanism for small object detection on drones is proposed.Firstly,the Transformer is introduced into the original YOLOv5 CSPDarknet backbone to improve network performance and obtain richer contextual information.Secondly,Convolution Block Attention Module(CBAM)and Spatial Pyramid Pooling-Fast(SPPF)are fused in the neck network to aggregate different receptive fields to capture multi-scale features in the image for improving object type and location detection performance.Finally,the shallow network is deepened based on the original network to construct a four-scale detector based on CSPDarknet-DUAV to improve the detection effect of dense small objects.The algorithm is applied to two object detection benchmark datasets,UAVDT(Unmanned Aerial Vehicle Detection and Tracking,UAVDT)and Vis Drone2021-DET(The Vision Meets Drone Object Detection in Image challenge 2021,Vis Drone2021-DET).The experimental results show that on the UAVDT dataset,the mAP(mean average precision,mAP)of YOLO-DUAV is as high as 34.2%,which is 3.0% higher than that of YOLOv5 s,and on the Vis Drone2021-DET dataset,the mAP of YOLO-DUAV is as high as 36.1%,which is 2.3% higher than that of YOLOv5 l.Compared with mainstream object detection algorithms such as Faster RCNN,SSD,Clus Det,and YOLO series in recent years,YOLO-DUAV can obtain the best indicators on these two datasets.Therefore,YOLO-DUAV has a high average precision and significantly improves object detection performance in challenging scenes such as complex environments at night,high altitudes,and foggy weather.(2)The nighttime small object detection of drones based on the Silicon-based golden light LED source is proposed.This article discusses the impact of different light sources on nighttime object detection from the perspective of the data source scenes.Firstly,a UAV-LED-G pedestrian dataset was collected and constructed Silicon-based golden light LED by using UAV.Next,nighttime images from the UAVDT and Vis Drone datasets were selected to construct UAVDT-night and Vis Drone-night datasets under general light sources.Finally,the YOLO-DUAV and YOLO series algorithms were applied to the three nighttime scene datasets.The experiments showed that on the UAV-LED-G dataset,the mAP of YOLO-DUAV reached 93.5%,while the YOLOv3,YOLOv4,and YOLOv5 x algorithms achieved 87.5%,89.3%,and 90.0%,respectively,which were much higher than the detection results under general light sources.In the experiments on the UAVDT-night and Vis Drone-night datasets,the mAPs of YOLO-DUAV were 76.4% and 52.5%,respectively,and the detection results in all three nighttime scenes were better than those of other YOLO series algorithms.Therefore,compared with general light sources,the Silicon-based golden light LED source can significantly improve the accuracy of small object detection in nighttime scenes,and the YOLO-DUAV algorithm has significant advantages in nighttime object detection tasks.(3)Edge computing-oriented UAV object detection algorithm YOLO-UAVLight is proposed.YOLO-UAVLight replaces the CSPDarknet network of the original YOLOv5 s with a ShuffleNetV2 network to reduce the model’s parameters and computational complexity,achieving a balance between speed and accuracy.YOLOUAVLight is applied to three datasets: UAVDT,Vis Drone,and UAV-LED-G,and compared with other lightweight algorithms.The algorithm is deployed on the Jetson Nano device,and real-time detection of objects from the UAV perspective is achieved.The experiments show that YOLO-UAVLight achieves the mAP of 87.3% on the UAVLED-G dataset,with model parameters and computational complexity of 3.68 M and8.1G,respectively,which are about 1/2,1/6,1/13,1/24 and 1/2,1/6,1/13,1/29 of the YOLOv5 s,YOLOv5m,YOLOv5 l,and YOLOv5 x models.Achieving an optimal balance between precision and speed.By harnessing the power of a camera for instantaneous detection,the algorithm boasts an impressive inference speed of up to 30 frames per second.This enables seamless real-time detection and makes it an ideal candidate for integration with drones,enhancing their capabilities for object identification and tracking.To sum up,this article delves into object recognition within the context of UAV,focusing on the aspects of algorithms,data,and end-user applications.It presents the innovative YOLO-DUAV algorithm,effectively tackling the challenge of subpar detection accuracy for small objects encountered by UAV.Additionally,the article unveils a tailor-made pedestrian dataset,UAV-LED-G,developed under the influence of the silicon-based golden light LED sources.This dataset significantly improves the detection precision of small objects during nighttime scenarios by employing cuttingedge illumination technology.Moreover,the article puts forward a streamlined algorithm,YOLO-UAVLight,which is implemented on the edge computing platform Jetson Nano.This approach results in impressive accuracy and rapid inference speeds,culminating in a real-time object detection system.