A Method Of Multi-level Pointcloud Data Segmentation And Typical Object Recognition

The acquisition of 3D semantic information of the scene is of great significance to the production practice in many fields such as surveying,mapping,remote sensing,navigation and crime scene protection.Semi-automatic semantic recognition through image data or manual measurement annotation is a commonly used method at this stage,but its degree of automation is low and consumes a lot of manpower and material resources.And data acquisition is greatly affected by objective factors such as time,light,weather,etc.,and the efficiency is low in practical applications.With the development of hardware such as depth cameras and lidar scanners,the acquisition technology of 3D point cloud information has become more mature,and the cost has gradually decreased.Obtaining 3D spatial information of the scene directly through the point cloud has been a widely used method.Compared with manual measurement annotation and identification through image data,object recognition through point cloud data has obvious advantages.First,its data acquisition is less affected by objective conditions,and the lidar scanner can work during the day and night.Secondly,the point cloud directly acquires the three-dimensional spatial coordinates of the surface of the target,without going through complex data processing steps such as motion recovery structures,which reduces the accuracy loss and information loss caused by the data processing process.In order to extract the information of the 3D laser point cloud accurately and efficiently,it is necessary to study the method of automatic classification and target recognition of the point cloud.For example,extracting ground point clouds from scene point clouds,that is,filtering,can reduce the amount of manual work in the process of generating DTM.Segmentation and extraction of point clouds and the identification of typical targets such as rods,roads,buildings,and power lines can reduce the amount of manual work and improve efficiency in the process of automatic driving high-precision map making and power inspection.In response to the above problems,this paper proposes a method of multi-level pointcloud data segmentation and typical object recognition.The specific research contents are as follows:1.A point cloud filtering method based on piecewise energy function optimization is proposed to obtain DTM elevation information.Firstly,handle outliers and rasterize the point cloud.Then the ground elevation of each grid is calculated by piecewise energy function minimization.The lowest point of each grid is a ground seed if its elevation is within the quantization error range.Finally the ground seeds are used to construct a Delaunay triangulated irregular network,the points that are closed enough to which are classified as ground points.The results of experiments on various terrain scenes proved that the total error of the filtering method is 3% lower than g Li DAR and it has the advantages of simple parameter setting and considering both accuracy and efficiency.This method can identify ground points more accurately and obtain more reliable ground elevation information.2.A method of multi-level pointcloud data segmentation and feature extraction is proposed.First calculate the local features of the point cloud,use these features as point node attributes to construct an energy function,and perform geometric consistency segmentation to obtain a simple structure with similar geometry,color and other attributes,called a super point.Then at the super-point level,the global features of the segmented objects are extracted.Then use the random forest classifier for training and classification to get the category inference of the segmented object.The algorithm is verified on the public point cloud data set 3DIS.The experimental results show that the algorithm makes full use of local and global information,and the target recognition accuracy is high.3.In the description of global features,the combination of shape features,general geometric features and local description statistical histogram features not only describes the local features of the object,but also reflects its global spatial distribution and geometric information.In addition,the classic local descriptor rotation image is improved to make it suitable for the application scenario of this paper and become a global feature.The results of classification experiments show that this feature can describe the target comprehensively and in detail.