Segmentation And Localization Of 3D Target Objects In Indoor Scenes

3D vision is a rapidly developing field in recent years,and it has become a major research focus in the field of computer vision.Many 3D vision sensors created in recent years have played a role in promoting research and engineering in various fields such as autonomous driving,industrial monitoring,virtual reality and target capture.3D visual data contains richer information than 2D images,before applying the information data contained to the actual use,the computer must first understand the 3D scene.Point cloud segmentation is a common form of understanding point clouds,which is of great significance for the realization of target object capture,point cloud space obstacle avoidance,path planning,AR reality and other follow-up research.However,the disorder and complexity of point clouds make it difficult to segment.The data point cloud scene has a large amount of data and a high degree of redundancy,the actual amount of useful target object information is very small.At the same time,more resources are required for storage and calculation,which leads to long calculation time for scene segmentation.In addition,different type of objects in the point cloud scene are not balanced,which increases the difficulty of network training,and many small goals are often overlooked.With the help of the more mature 2D image segmentation technology with longer history,this paper designs and proposes an indoor scene target object positioning and segmentation system with point cloud processing and neural network-based methods.3D color point cloud data cannot be obtained directly from RGBD cameras,but are constructed from depth maps and RGB images.In order to solve the problem of poor detection of small objects,this article first uses the YOLOv5 network to detect small targets on the color picture quickly and accurately.The framed area is mapped into the three-dimensional space to enhance the correlation between the image and the point cloud.The subsequent segmentation of small objects will be performed separately in these processed candidate areas.In order to separate the target object from the background in the point cloud area,this paper designs two point cloud segmentation systems,which respectively use traditional methods and deep learning methods to segment the point clouds and compare them.The traditional method uses the RANSAC fitting method and a region growth method with color to remove the background point clouds;the other method designs a neural network to separate the target object from the background.The network is based on the PointNet network and has undergone many optimizations.The improvement in the network model is mainly reflected in the following points:First:In order to better extract the features,an autoencoder network is used to replace the global pooling in the original method when extracting the global features,and it is better to retain global network characteristics.Second:A unique loss function is designed for the unique network of this article,which ensures the balance of category data and at the same time obtains the integrity of the global feature information.Third:increase an attention module to enhance the feature extraction ability of the network.We use this network to classify the ShapeNet data and the results obtained a better segmentation accuracy than the original network.In order to compare the performance of the two segmentation methods,we analyze the results of different detection difficulties on the RGB-D Scenes dataset to test and verify its feasibility and stability.Finally use the deep learning segmentation methods for the segmentation of the small point cloud of the system in this article.For the segmented target object,this paper uses the oriented bounding box based on principal component analysis to express its location.At last,the overall experiment is carried out on the RGB-D Scenes dataset and the real indoor complex scene,which proves the applicability and robustness of the experiment.