Point Cloud Target Recognition And Matching Based On Deep Learning

Deep learning and SLAM is now two important research directions in the field of computer vision,among which convolutional neural network in deep learning can realize the semantic understanding of image by imitating biological brain and visual mechanism;In addition,SLAM can realize the tracking through camera principle and mathematical principle,and also can realize the point cloud mapping.With the continuous improvement of information technology and hardware performance,the application of deep learning and SLAM is becoming more and more extensive.This paper studies how to integrate them to promote each other.In this paper,from the theoretical research of deep learning and SLAM to their integration will be introduced systematically,the semantic segmentation and optimization of positioning and semantic point cloud construction are combined into a simple semantic SLAM system.It solves the problems of poor accuracy of slam in dynamic scene and incomprehensibility of point cloud mapResearch on the method of image semantic segmentation based on deep learning.The principle of convolution neural network,the structure of existing semantic segmentation network and loss function are described in detail.On this basis,a joint segmentation loss function combining pixel loss and mask loss and a multi-scale semantic segmentation convolution neural network which can enhance the receptive field are proposed.Research on the method of localization based on stereo camera,in which the imaging model and motion model of the camera are described in detail,and the method of solving camera pose and how to estimate camera motion path are introduced.Research on the matching method and filtering method of feature points,and proposes a dynamic scene disturbance reduction method based on semantic segmentation,which can effectively reduce the disturbance of pose estimation in dynamic environment,and improve the accuracy of camera localization and mappingResearch on the semantic point cloud construction method based on semantic segmentation,and establishes the semantic point cloud map for the space where the camera moves by combining the semantic mask of the image,so as to achieve the purpose of point cloud recognition,and provide the precondition for the machine to understand the scene and interact with it.