Indoor Object Detection And Pose Estimation Based On Convolutional Neural Network

Object detection and pose estimation from a single image have always been very an important research task in the field of computer vision.With the advent of deep learning,the advantages of convolutional neural networks in the field of object detection and pose estimation are becoming more and more apparently.Most of the current pose estimation algorithms use direct or indirect regression prediction methods.However,some pose estimation algorithms can not realize the function of object detection and multi-target pose estimation,and the accuracy of object pose is lower.Some pose estimation algorithms calculate the object pose by using the key-points of the object and camera parameters,its accuracy is higher than the direct prediction methods,but more postprocess is needed.In this paper,object detection and pose estimation are combined together based on convolutional neural network to realize multi-target object detection and 6D pose estimation,and the accuracy of object detection and pose estimation of direct prediction methods and indirect prediction methods are improved,which solves the problem that some algorithms cannot carry out multi-target analysis.The main work of this paper is as follows:Firstly,we propose a multi-object object detection and pose estimation algorithm based on convolutional neural network(CNN).The algorithm improves the object detection network Faster R-CNN,realizes end-to-end object detection and pose estimation,and utilizes a full convolution network(FCN)replaces the full connection layer,greatly reduces network's parameters,and improves the downsampling method after the RPN network to improve the downsampling accuracy.The algorithm improves the accuracy of object detection and can handle the problem of multi-object estimation.The classification and regression methods are used to directly estimate the pose of the object,which greatly improves the accuracy of the direct prediction methods.Secondly,based on Faster R-CNN network backbone,multi-object object detection and pose estimation can be achieved by combining the vertex of object 3D bounding box and PnP algorithm.The algorithm transforms the problem of pose estimation into the problem of coordinate prediction of key-points of objects,and estimates the pose of objects through coordinate points and PnP algorithm.Finally,we integrate improved hourglass network and the latest object detection network to estimate the pose of the object by predicting the heatmap of the eight vertices of the object 3D bounding box.The coordinate points prediction problem is transformed into the coordinate points' heatmap prediction problem,thereby improving the accuracy of the coordinate points prediction and further improving the accuracy of the pose estimation.