Research On 3D Object Detection For A Sweeping Robot Based On Image Voting And Contextual Information Enhancement

Most of the current 3D object detection methods are limited to 2D object detection,which requires converting 3D point cloud data into regular grids(such as voxel grids or bird ’s-eye view images)or extracting 3D frames through detection results in 2D images.However,there are fewer ways to detect 3D objects directly using 3D point cloud data than using 2D detectors to detect 3D objects.This is because the center of mass of a 3D object may be far from any surface point,making it difficult to accurately return in one step.To solve this problem,we study Votenet,an end-to-end 3D object detection network based on deep point set network and Hough voting.However,Votenet has its inherent limitations,as it only uses point cloud data,which is sparse and lacks color information,and is susceptible to sensor noise.In contrast,the images have high resolution and rich textures that can provide supplementary information about the 3D geometry of the point cloud.In addition,Votenet only focuses on the identification of a single object,ignoring the context information between objects.Therefore,we consider using RGB images and context information to enhance Votenet to improve recognition accuracy.Specifically,in this method,we investigate a 3D object detection structure specifically for RGBD scenes based on the fusion of 2D and 3D voting.We extract geometric and semantic features from2 D images and use camera parameters to lift these features into 3D space.At the same time,we used a patch-to-patch context(PPC)module to capture the context information between point patches,and introduced an object-to-object context(OOC)module to capture the context information between candidate objects before the proposal and classification stages of candidate objects.In addition,we designed the Global Scene Context(GSC)module to learn global scene information.Our approach can improve the accuracy and robustness of 3D object detection by simultaneously capturing contextual information at the patch,object,and scene levels.Experimental results show that m AP(mean average accuracy)is 5.7% higher than Votenet when IOU is 0.25.With the continuous development of The Times and the continuous improvement of technology,people’s life is becoming better and better.Large-scale sweeping robots have become essential products in shopping malls and transportation hubs.However,at present,there are many manual sweeping robots on the market,which need to spend a lot of time and waste manpower when cleaning the whole large area.For this reason,we developed an intelligent sweeping robot based on the method in this paper.Since binocular cameras are used instead of multi-line radars,and binocular cameras are cheaper than multi-line radars that can achieve the same effect,this method has strong practicability based on the consideration of cost performance.