Research On Overlapping Target Detection On Pipeline Based On YOLOv3

Object detection is a popular field of computer vision research,and is also the key to object tracking,gesture recognition and behavior recognition.It is widely used in video object intelligent monitoring,object behavior gesture recognition,pipeline product detection recognition,medical image analysis and other systems.When the pipeline products are overlapped or blocked,the product images captured by the camera cannot be recognized as upper and lower objects by the robot arm,resulting in the robot arm not grasping,erroneously grasping,or crashing,resulting in disordered product packaging and assembly line products Backlog,This paper analyzes and studies the optical flow method,the inter-frame difference method and the background subtraction in the traditional moving object detection algorithm based on the gray pixel of the statistical background model,and several mainstream neural network frameworks based on deep learning,And deep learning network architecture,extract the upper overlapping object of the overlapping products in the pipeline,the main work is as follows:(1)Introduce the research background and significance of object detection,the current research status and the basic theory of digital image processing,research on the optical flow method,the inter-frame difference method and the background subtraction method in the traditional object detection algorithm based on the pixel gray statistical background model.The basic theories related to deep learning and feature extraction classification network structure are analyzed,and several mainstream network structures based on deep learning are compared and analyzed.This paper proposes an improved background modeling of adaptive Gaussian mixture model based on statistical background model and YOLOv3 architecture combined with improved algorithm of residual dense network based on deep learning neural network.(2)Starting from the background separation algorithm before Gaussian background modeling,this paper establishes a background model for images containing moving objects in the video sequence,resets the learning rate changes in the parameter update strategy and the parameters of gaussian distribution,and improves The background modeling algorithm flow of Gaussian mixture model based on statistical background model is introduced.At the same time,in the preprocessing of video sequence images,the denoising filter kernel selection,threshold segmentation threshold selection,and object morphology filter kernel selection are performed on the image containing the detection object.And comparison of edge detection operator selection.Experiments show that the improved adaptive Gaussian mixture model background subtraction algorithm has significantly improved the running time and detection accuracy.(3)Starting from the network structure of the YOLOv3 algorithm,this article combines the residual dense blocks to perform feature clarification,feature extraction and upsampling operations on the low-resolution feature maps,and then uses the residual dense network to analyze the before and after the residual dense blocks.The local features and global features are fused,and after the feature classification is performed by the convolution output layer and the Softmax classifier,the recognized feature attributes are output.Perform sample-label and cluster sample-object preselection on the sample training set,and change the attribute-configuration files of the network structure to train and test the labeled-samples dataset.Experiments show that the improved YOLOv3 combined with RDB network structure has significantly improved the running time and object detection accuracy,and it has good robustness to overlapping object detection.