Research On Three Dimensional Object Recognition Algorithm For Self-driving In Typical Scenarios

This topic is based on the perception of the environment of the driverless vehicle.Based on the laser point cloud data,the identification of people and cars in a typical environment is studied.The research content is mainly divided into thre e parts: preprocessing of point cloud,target recognition based on local features,and target recognition based on 2d features.The main work of the dissertation is as follows:First,in the preprocessing part of the point cloud,a combination filtering method is proposed for the point cloud data with poor quality of the indoor corridor that is actually collected,and a better processing effect is obtained.In order to compare and analyze the original data and the data processed by different methods,several point cloud quality evaluation indicators were put forward and the results were quantitatively analyzed.Second,the target recognition based on local features is studied.Because there is no suitable data set for algorithm validation,the use of Blender software to create model and field-of-view cloud datasets is proposed.Some of the theories such as feature extraction,feature description,feature matching and recognition have been well understood and studied.For the problem of insufficient descriptio n of a single descriptor,the idea of descriptor fusion is proposed.Simulation results show that the proposed method has certain rationality and effectiveness.Finally,the 3D target recognition based on two-dimensional features is studied,and the 3D point cloud is projected onto a 2D plane to extract two-dimensional features for target recognition.First,the three-dimensional point cloud data is projected to a two-dimensional space;Secondly,the projected point cloud is binarized in a two-dimensional space;Then,the feature descriptors of the point cloud are described using the descriptors of the two-dimensional image.Finally,the experimental results show that it is feasible to use the two-dimensional features of point clouds for 3D target classification and recognition.