Research And Application Of Multi Scene Object Detection Method Based On YOLOv5

Object detection is one of the research directions in computer vision,and deep learning based object detection technology is becoming increasingly mature in aviation,industry,and natural scenes.In these scenarios,deep learning based object detection methods may face problems such as complex image backgrounds,large aspect ratios of objects,drastic changes in aspect ratios of objects,and uneven distribution of objects,which can degrade the performance of object detection.To address the above issues,this article conducts research on object detection methods based on the existing cutting-edge object detection method YOLOv5.The main work is as follows:(1)A rotating object detection method YOLOv5-CBCA is proposed.A network model incorporating convolutional channel attention is designed to enhance the semantic and location features conveyed to the top and bottom layers of the feature pyramid;optimize the object box regression function of YOLOv5 to adaptively generate a wide and high regression range;design an object box loss function that includes coverage area,center point distance,aspect ratio,and angle loss,demonstrating the completeness of the loss function and meeting the basic conditions of the loss function.In order to verify the effectiveness of the method proposed in this paper,experiments are conducted on remote sensing image UCAS-AOD dataset,HRSC2016 dataset,and industrial scene QR dataset.The experimental results show that when using m AP and FPS as evaluation indicators,the method in this paper has a certain improvement in performance compared to some representative methods in recent years,and the method can effectively achieve two-dimensional code positioning for industrial scenes.(2)An improved object detection method based on k-means is proposed.An anchor box clustering algorithm based on gaps is designed.After clustering,the spacing between clusters is increased,so that the anchor boxes do not interfere with each other when predicting the object box,and the prediction range covers more objects,especially those distributed on the edge with fewer numbers.An anchor box clustering algorithm based on regions is designed.During the first clustering,the cluster is refined to achieve local optimization.Secondary clustering with local optimal results makes the clustering results closer to the global distribution.Anchor boxes pay more attention to the prediction of all objects on the global level.In order to verify the effectiveness of the proposed method,experiments are conducted on remote sensing image UCAS-AOD dataset,natural scene COCO dataset,and industrial scene QR dataset.Experimental results show that when using m AP and FPS as evaluation indicators,the method in this paper has a certain improvement in performance compared to some representative methods in recent years,and the method can be effectively applied to multi scene object detection.(3)An object detection system based on YOLOv5 is designed and implemented.In order to verify the practicality of the above methods,the industrial QR code data set QR and fatigue driving data set TIRED are produced.The industrial QR code detection system and fatigue driving detection system is designed and implemented on the NVIDIA Jetson nano development board,achieving the detection of industrial QR codes and driver status.