Research On Wheel Object Detection In 3D Point Cloud Oriented To Intelligent Garage

Li DAR has been widely used in many fields as an effective means of spatial information acquisition.As one of the main technologies in the application of Li DAR,3D point cloud data processing has gradually become an important research field.This paper focuses on data processing of point cloud in intelligent garage.Aiming at the requirement of real-time acquisition of wheel position and attitude information for parking vehicles in intelligent garage,this paper proposes and implements a 3D point cloud wheel object detection method based on Deep Convolutional Neural Network according to the characteristics of wheel object in point cloud data.(1)Considering the one-dimensional characteristics of wheel centerline and the real-time requirement of the system,3D point cloud is transformed into 2D sequence inputs.Using recurrent neural networks,the position of a wheel centerline is located by foreground/background probability of the serialized input in each step of the time series.In this paper,the recognition effects of recurrent neural networks,long short-term memory networks and gated recurrent unit networks are compared and a wheel center positioning network based on its variant long short-term memory networks is designed.The experimental results verify the effectiveness of the application of recurrent neural network in wheel centerline positioning.(2)To solve the problem of wrong detection caused by the large positioning deviation when the wheel is occluded,a parallel fusion of recurrent neural networks and Faster R-CNN called Fusion RNN is proposed in this paper.With the processing time series ability of RNN and the ability of extracting the spatial hidden features of CNN,Fusion RNN can analyze the semantic relationship between serialized vectors and identify the geometric characteristics of the wheel.(3)To obtain wheel attitude information for automatic parking,a new point cloud object detection model,Loc Point Net is proposed in this paper.Location module is added on the basis of the classification network,Point Net,and regional proposals are introduced to regress the location information of the wheel.Loc Point Net solves the problem that 3D point cloud object detection models based on voxels are not able to optimally predict the location and attitude of the object by using bounding boxes along voxel directions in small scenes.Loc Point Net uses the minimum bounding box to represent the object,which not only approximates the real contour,but contains attitude information of the point cloud object as well.To sum up,in the intelligent garage system,using the information contained in 3D point cloud collected by lidar,this paper applies deep learning in wheel object detection,locating the vehicle wheel center,recognizing the wheel attitude.The experimental results on Wheel Pro 2D vehicle image data set show that bidirectional long short-term memory networks have the best comprehensive positioning effect among several recurrent neural networks with fast positioning speed;Fusion RNN effectively suppresses the wrong detection phenomenon caused by wheels which are partially occluded by obstacles.The experimental results on Wheel Pro 3D vehicle point cloud data set prove that Loc Point Net has the ability to dectect wheel objects.In particular,the prediction of wheel attitude is more accurate.