Research On Indoor Object Detection Based On 3D Point Cloud

In recent years,artificial intelligence related products can be seen everywhere in our lives.Intelligent robots can complete many operations,thereby freeing people's hands,which is the future development trend.The threedimensional point cloud contains the spatial position information and the geometric structure information of the object,which has a natural advantage over the two-dimensional image.In recent years,point cloud related research has become a hot direction in the field of artificial intelligence.Point cloud indoor target detection is a technology for indoor robots to analyze surrounding environment information,which is equivalent to giving indoor robots "eyes" to see objects clearly,which has important research significance.This paper studies the indoor object detection in 3D point cloud,and the research content is as follows:First of all,the existing 3D lidar scans at a fixed resolution.In an unknown scene,collecting data with a fixed resolution scanning method may not be able to capture detailed information for complex objects.For simple objects,using too high resolution will cause data redundancy.Based on the two-dimensional lidar and rotating gimbal,this paper builds a three-dimensional point cloud data acquisition platform that can scan with different resolutions according to the complexity of the objects in the scene.It strikes a balance between capturing complex structures and reducing the amount of data.Secondly,there are many current researches on point cloud object detection,but most of these researches are applicable to outdoor environments.The difficulty of detecting indoor objects relative to outdoor objects is that they are smaller in size and more detailed.The point cloud data is sparse,which is not conducive to detecting small objects.In order to improve the detection accuracy of indoor objects,an improved Vote Net network based on semantic features is proposed.First,the model adds a pre-network that can effectively extract semantic features.By fusing the extracted semantic features with geometric features,more effective feature blocks are obtained for target detection.Secondly,by adding semantic constraints in the generation stage of Vote Net voting clusters,the problem of semantic inconsistency in the midpoints of Vote Net voting clusters is solved to a certain extent.Experiments are conducted on the public data set Scan Net,and the experiment proves that the improved network proposed in this paper has higher detection accuracy than Vote Net when detecting indoor objects.Finally,in order to achieve a complete point cloud indoor object detection system,the Vote Net improved network based on semantic features is used to detect the object on the point cloud data obtained by the point cloud scanning platform in this paper.Experiments show that the improved target detection network proposed in this paper can detect objects from the point cloud data obtained by the point cloud scanning platform of this paper.