Method For Object Classification And Detection Based On Multi-view Images Of Outdoor 3D Point Cloud

With the development of 3D scanning technology,3D point cloud data can be obtained quickly by using point cloud vehicle mounted scanning device or moving point cloud scanning device,which provide the data base for object classification and object detection in 3D scene.With the availability of hardware and the expansion of large-scale data sets,the methods such as deep learning can be applied to the complex 3D data.These techniques provide hardware and software conditions for the research work in 3D scene.Objects classification and detection is a fundamental research area in outdoor three-dimension(3D)point cloud scene,this technology can be used to support the automatic driving,urban road planning,3D map navigation and so on.At home and abroad the data of three-dimensional on classification and detection is mostly used simple data in door,rarely directly use the data on the large-scale outdoor.3D point cloud because of its huge amount of data and information redundancy,direct use of learning methods to deal with it is difficult to achieve good results.In this paper,two solutions are proposed based on image mapping and deep learning in point cloud data:one is to classify the single-point point cloud tree species and the other is the vehicle detection under the point cloud scene.The main work and innovations of this paper include:The main work of the paper mainly includes two parts:The one part is tree classification in single target of point cloud.A set of overall process are put forward include point cloud tree extraction,data preprocessing,3D point cloud rasterization,and deep learning classification.And Deep Belief Network(DBN)is used for fitting the small-scale data.In the training process,3D point cloud rasterization is used to generate low-level feature as input of deep learning model.The multi-angle criterion is adopted to improve the classification accuracy.The other part is vehicle detection in point cloud scene.Mapping relation of point cloud and multi-view images is used to constrain the location of the vehicle in the original point cloud scene.First,model is trained on the public data set,and then fine tuning is carried out in the data scenario of this paper.In order to solve the multi-scale problem of vehicle size in the data,the convolutional neural network is modified for detecting multi-scale targets.These works improve the detection accuracy.