3D Object Detection In RGB-D Scenes Based On Deep Learning

As an important issue in graphics and 3D vision,3D object detection has received widespread attention from both industry and academia.Its purpose is to detect specific3 D target objects in a given scene and determine their position and category information.In indoor scene 3D object detection based on RGB-D data or discrete point cloud data input,due to the varying appearance,shape/pose,and various factors such as object clutter and mutual occlusion commonly present in the indoor environment,it is difficult to effectively utilize scene data information and low object detection efficiency,making indoor scene 3D object detection a challenging topic.Therefore,for the problem of 3D object detection in indoor scene RGB-D data,how to effectively extract and fully utilize scene object feature information,while excluding background noise interference,and achieve fast and accurate object detection in complex indoor scenes is a challenge.However,scene feature information extraction based on neural networks,effective separation of scene foreground and background,and effective sampling of scene point clouds overcome the limited utilization of data information in 3D object detection One of the effective methods for solving problems such as high noise interference and low detection efficiency.The main work and research results of this article are as follows:(1)In order to overcome the difficulties of 3D object detection in RGB-D scenes,such as poor adaptability to complex indoor backgrounds,and difficulty in effectively utilizing object region information and scene point cloud feature information during object detection,based on the guidance of object region information,this paper proposes a 3D object detection framework that integrates global and local point cloud features and eliminates background interference.This framework takes scene RGB-D data as input,first expanding the 2D region of the target object into 3D oblique cone point cloud data;Then,the global and local features of the point cloud are fused,and the probability score of the correlation between each sampling point and the foreground background is predicted.Based on this score,scene background points are removed to form a shielding point cloud;Finally,3D object detection is performed in a shielded point cloud and recommendations are made based on object area information.(2)In order to shorten the inference time of 3D object detection models and ensure the accuracy of object detection,a lightweight RGB-D scene 3D object detection method based on fusion sampling strategy is proposed.This method takes scene RGB-D data as input,and first converts it into a 3D point cloud through a two-dimensional convolutional neural network and a depth camera projection matrix;Then,a fusion sampling strategy is used to sample the points in the scene,and feature representative points are preserved through fusion sampling;Finally,a deep Hough voting mechanism is used in the scene feature representative point cloud to vote on the center of each object in the scene,and relevant features around each object are clustered to achieve 3D object detection in the scene.