Vehicle Simultaneous Detection And Motion Prediction Based On Continuous Lidar Point Cloud Sequence

In recent years,with the development and maturity of artificial intelligence technology,the application of artificial intelligence technology has become more and more extensive,from image data to voice,text,laser point cloud and other data.Autonomous driving technology needs to process various types of data,so it has become an important application scenario for artificial intelligence technology.Autonomous driving technology consists of four parts: environment perception,behavior decision,path planning and motion control.This dissertation mainly studies vehicle object detection under environmental perception and vehicle motion prediction under behavior decision,and the object detection and motion prediction are performed at the same time instead of the object detection first and motion prediction second.In this dissertation,the design of simultaneous vehicle detection and motion prediction method based on continuous frame laser point cloud mainly includes five research contents: dynamic rasterization,raster feature encoder,object detection network,multi-frame fusion module,and motion prediction module.Rasterization and raster feature encoders are common methods for converting laser point cloud data into structured data similar to images.Aiming at the shortcomings of the existing rasterization methods that occupy a large amount of memory,lose the original point cloud information and need to define various thresholds in advance,this dissertation proposes a new rasterization method,namely dynamic rasterization.Aiming at the different advantages and disadvantages of the existing two commonly used raster feature encoders,this dissertation proposes a new raster feature encoder,which combines the advantages of the two,and the advantages are high speed and fine feature.Compared with most existing object detection networks that need to define anchor,this dissertation uses a fully convolutional neural network to design a object detection algorithm that does not require anchor,and directly predicts the bounding box of the object on the network output feature map.Based on the existing methods,using three-dimensional convolution or convolutional recurrent neural network to design a multi-frame fusion module to extract sequence motion information,this dissertation proposes to design a multi-frame fusion module using three-dimensional convolution and time series maximum pooling.In order to make full use of label information,this dissertation proposes a motion prediction module,which predicts the motion vector representing the object's past and future movement,and sends it to the recurrent neural network to further improve the performance of motion prediction.On the public dataset,the method in this dissertation has improvement in object detection performance and motion prediction performance compared with existing methods.