Research On Autonomous Driving Trajectory And Behavior Perception Based On Graph Convolution

Automated Driving Systems are complex system that integrates control technology,sensing algorithm,path planning,spatial modeling and positioning,etc.Over the past decade,with the development of deep learning technology and the improvement of computer performance,autonomous driving-related technologies have also developed rapidly.But there are many obstacles to widespread self-driving technology today,such as ensuring adequate safety.Among them,vehicle trajectory prediction and its behavior analysis are key technology,which provides necessary information for vehicle safe driving assistance,path planning,decision-making,and safety warning of surrounding environment.For track prediction and driving behavior analysis,from solving insufficient time and spa ce depend on the feature extraction,dynamic driving scene features are difficult to extracting,based on vehicle driving behavior problems such as interactive background analysis,this paper carried out the following research,contains the relevant data construction,track prediction,behavior analysis three priorities:1)A semi-global graph generation algorithm is proposed,which solves the problem of matrix sparsity when the graph model describes the driving scene,and the problem that the time dimension matrix has no correspondence relation.A graph model is introduced to model the driving scene.The vehicles in a driving scene are modeled as nodes of the graph,and the relationship between vehicles is represented by the weighted edges of the graph.On this basis,the concept of dynamic graph is introduced to describe the dynamic driving scene.2)A vehicle trajectory prediction model based on graph convolution is proposed.The model consists of two types of graph convolution networks based on spectrum convolution and a enc-dec network.For vehicle trajectory prediction task,in order to perceive the interaction between vehicles,a graph convolutional network is introduced.The graph convolution based on spectral theory can solve the problem by mapping the Laplace matrix from the spatial domain to the frequency domain by Fourier transform.The Laplace matrix formed by the normalization of the adjacency matrix can describe the information potential difference between nodes.This kind of neural network can well solve the problem of extracting the features of the relationship between vehicles.In addition,the dynamic graph describes the driving scene with dynamic changes.In order to efficiently extract the dynamic features in the dynamic graph,we introduce the new product method M-Product,and build a variant form of the graph convolution neural network based on the M-Product operation,so that the graph convolution operation can run directly on the three-dimensional tensor.The problem of extracting dynamic features from dynamic graphs is solved.3)A driving behavior analysis algorithm based on graph model is proposed.In order to consider the interaction between vehicles in the analysis of driving behavior,based on the aforementioned working background,a graph model is introduced to shift the focus of the characteristics of the target vehicle to the differences with surrounding vehicles,so as to better describe the behavior of a vehicle.At the same time,the centrality function and other evaluation indexes were introduced to describe the topological characteristics of the graph model,and the feature functions describing different driving behaviors were constructed to classify vehicle behaviors.