3D Reconstruction Of Zebra Crossings Based On Procedural Plane Fitting From Point Clouds

Zebra crossings,as a common road marking,are widely used in the world.It is usually composed of many parallel rectangular planes,which provide important traffic guidance for drivers and other road users.The zebra crossings model and the geometric information can be obtained by the three-dimensional(3D)reconstruction of zebra crossings from point clouds,which effectively solves the problem of large storage in high-precision maps,low efficiency and high cost by manual measurement and uneven distribution of point density.Therefore,it is of great significance to study the 3D reconstruction method of zebra crossings from point clouds.The main contribution of this thesis is following the two aspects.Firstly,a zebra crossings 3D reconstruction method based on cuckoo search algorithm is proposed.3D reconstruction of zebra crossings from point clouds can be transformed into the plane fitting problem of zebra crossings.To solve the problem of incomplete geometric information of the plane model obtained by the current point clouds plane fitting method,this thesis adopts the procedural plane fitting method and designs a suitable probability program to obtain a plane model with more complete geometric information.The cuckoo search algorithm is an optimization algorithm for procedural plane fitting to perform the 3D reconstruction of zebra crossings.Our zebra crossings model reconstructed not only has boundary information,but also can obtain geometric information such as the length,width and slope of the zebra crossings.Secondly,we proposed a zebra crossings 3D reconstruction method based on deep reinforcement learning algorithm.This method realizes multi-plane fitting,and uses deep reinforcement learning with strong decision-making ability as the optimization algorithm of the procedural plane fitting to perform 3D reconstruction of zebra crossings,which effectively reduces the number of iterations fitting.In this thesis,the two research contents are verified by experiments on simulated point cloud data and raw point cloud data,and the experimental results are analyzed.The experimental results show that our method is effective.