Pedestrian Object Detection In Forest Scenes Based On UAV Images

In recent years,with the rapid development of UAV remote sensing technologies,detection techniques based on UAV remote sensing images have been widely used in various research fields.Drones mostly work at high altitudes,and while high-altitude photography can cover as much of the detection object as possible,it also causes the object to be too small in the UAV image,which is not favorable for object detection algorithms.The high pixel resolution and rich image information also make the task of UAV remote sensing image object detection more difficult.Currently,most networks use downsampling to compress images when causes the loss of feature information.Therefore,how to efficiently identify detection objects in UAV remote sensing images with limited feature information has become a hot topic of research in various fields.In this thesis,we carried out extensive research in forest scenes based on UAV images using forest pedestrian video data from Chongli District,Hebei Province.1.To address the problem of reduced detection speed due to adaptive segmentation detection strategy,a high-resolution lightweight YOLO target detection network is proposed,which is improved based on YOLOv7.The network introduces a new loss function(GCKLoss)that reduces the sensitivity of small UAV pedestrian objects to position deviation.A receptive field block module(RFB-Attention)with enhanced attention mechanism is added to the backbone network.An adaptive segmentation detection strategy is proposed,which can retain as many features of the detected object as possible and significantly improve the accuracy for small object detection tasks.Finally,experiments are conducted on Chongli UAV pedestrian detection dataset to verify the effectiveness of the proposed method.2.The high-resolution lightweight YOLO target detection network is proposed for the problem of detection speed reduction due to adaptive segmentation detection strategy,which is improved based on YOLOv7.Aiming at the characteristics of pedestrian objects of UAV images and the defects of pyramid network,the YOLOv7 object detection network is improved by combining the high-resolution network(HRNet)and the lightweight module to reduce the number of parameters of the network.To further improve the algorithm accuracy,this thesis proposes a SMCP data enhancement method to optimize the algorithm operation results.The proposed high-resolution lightweight YOLO object detection algorithm reduces the number of parameters by 36% and improves the detection speed compared to YOLOv7.3.Combining the advantages of enhanced feature extraction network and high-resolution lightweight YOLO target detection network,a high-resolution enhanced feature extraction network structure is proposed,and the detection model is deployed on a server to build a visualization interface to complete the practical application of pedestrian detection tasks for UAV remote sensing images in forest areas.