Research And Application Of Object Detection Algorithm For Dense Scenes

As one of the basic functional materials,rebars are widely used in the construction industry.In the process of using and trading rebars,it is necessary to accurately count the quantity of rebars to avoid economic risks and disputes in subsequent handover links.The traditional method uses manual counting.This method is cumbersome,time-consuming and laborious,and the counting accuracy is affected by the number of rebars and the fatigue status of the inspectors.Therefore,enterprises have a great demand for improving the efficiency of rebar counting.At present,the general object detection technology based on deep learning performs well and is widely used in actual scenes such as traffic sign detection and face detection,but its application effect on the task of rebar counting in actual scenes is not good,mainly because the environment background such as weather and lighting in the actual scene is complicated,the scale and the quantity of the rebars are large,the layout of rebars is wide and the placement of rebars is uneven,and some rebars are uneven and densely covered by other rebars.Based on business needs and the feasibility of deep learning technology,this paper conducts an in-depth study on the task of rebar counting in large-scale dense stacking scenes,and proposed a rebar counting algorithm and multi-image stitching method for large-scale dense stacking scenes.Finally,the task of rebar detection and counting in large-scale dense stacking scenes is completed efficiently and accurately.The main work and innovations of this paper are as follows:1.Aiming at the characteristics of rebar detection and counting tasks in densely stacked scenes,a deep learning-based rebar counting algorithm called Rebar RCNN is proposed.Rebar RCNN is improved based on the two-stage detection algorithm Faster RCNN.The improvements mainly include network structure,loss function and postprocessing algorithms.The precision and recall of the Rebar RCNN model on the private test set are both above 98%,which is about 3% higher than Faster RCNN,and it is more robust to environmental changes such as illumination and weather.2.A multi-camera photo detection and counting scheme is proposed based on the problems of incomplete single-camera photography,blurred images of small rebars,and severe occlusion of rebars in large-scale scenes.And a multi-image stitching method was proposed to solve the problem of repeated counting of overlapping areas of adjacent images in the multi-camera photo detection and counting scheme.The multi-image stitching method based on the invariance of the relative position of the rebar bundle completes the statistics of the total number of rebars in the indoor experimental scene,and verifies the feasibility of the multi-camera shooting detection and counting scheme.The multi-image stitching method based on the invariance of the topological structure of the rebar distribution takes advantage of the fact that the relative position of the rebar in adjacent photos is basically unchanged,and completes the statistics of the total number of rebars in a large-scale scene.3.Designed and developed a complete set of intelligent rebar detection and counting system,and applied the above-mentioned rebar counting algorithm and multi-image stitching method to the ground.The system is easy to operate and has a beautiful interface.It provides functions such as image collection,file import,rebar detection and counting,manual verification,and result storage.It can perform real-time detection and counting of rebar pictures imported by users or collected by cameras,and allows users to modify and improve the detection results.