Research On Small Object Detection Based On Urban Traffic Image

With the development of science and technology,object detection is widely used as the basic algorithm of computer vision.There are a large number of small objects in real traffic scenes,and small object detection is a difficult problem in object detection.Therefore,it is of great significance to study small object detection in traffic scenes.This paper first analyzes the feature extraction network,designs a small object feature enhancement module and a variance-guided Region-of-Interest(Ro I)feature fusion module,and improves the detection accuracy of small objects by enhancing the small object features.Then explore the impact of the loss function on small object detection and the optimization of intersection and union is proposed respectively,and a penalty mechanism based on corner distance is set up.Finally,a vehicle cut in line recognition model based on object detection is proposed,and the object detection technology based on deep learning is applied to the field of urban traffic.The main innovations are as follows:1.A feature enhanced feature pyramid network is proposed.A small object feature enhancement module is proposed and embedded in a specific lateral connection in the feature pyramid to enhance small object features.A variance-guided Ro I feature fusion method is proposed,which weights and fuses multiple features through variance to enhance the features of Ro I.2.A loss function and penalty mechanism based on the intersection and union ratio are proposed.The intersection and union in the intersection and union ratio are optimized respectively,which speeds up the regression of small object candidate boxes.The corner distance is used as a penalty mechanism,and the standardization of penalty mechanism is discussed according to specific situation,which optimizes the regression effect.The experimental results show that the this function improves the detection accuracy and precision of small objects.3.A recognition model of vehicle cut in line at urban intersections based on object detection is proposed.The result of object detection is filtered by category,and the vehicle object is retained.Hough transform is performed on the center points of the vehicle object to fit the center lines of the lane.The distance between the vehicle and the center lines of the lane is used to detect whether the vehicle is cut in line.