Construction And Localization Of Multi-resolution Map Based On Vision

Map construction and localization is one of the key technologies in the unmanned vehicle driving research.Constructing high precision maps based on lidar and other sensors is the mainstream method at present.However,limited by the performance and cost of sensors,coupled with the complexity of the unmanned vehicle driving environment,high precision maps have many disadvantages,such as difficult to construct,high maintenance and use costs,and poor robustness to environmental changes.In order to solve these problems,this paper focuses on the research of vision-based multi-resolution map construction and localization technology,including feature extraction method based on convolution neural network,multi-resolution map construction and localization based on Non-uniformly Distributed Map Cell(NuDM-Cell),and multi-resolution hierarchical map multi-lane oriented construction and location method.The main works of this paper are as follows:1)Aiming at the application requirements of multi-resolution map construction and visual localization,the existing image feature extraction methods are studied and analyzed.Based on the separability of features and clustering effect,the performance comparison experiments are carried out on the "SIFT+BoW" histogram features,HOG features,GIST features and the features extracted by the convolution neural network.Experimental results show that the image features extracted by convolution neural network contain more abstract and comprehensive environmental information,and can be used for construction and localization of multi-resolution maps.2)A vision based construction and localization method for multi-resolution map is proposed and implemented.Based on the experience of human navigation and positioning,the concept of Non-uniformly Distributed Map Cell(NuDM-Cell)is introduced firstly,then the location relationship of sequential images is used as a prior constraint information,and the convolution neural network is used together with clustering algorithm to generate NuDM-Cell.Then a vision based multi-resolution map consist of NuDM-Cell is constructed.A locator based on convolution neural network is trained to realize real-time visual localization for NuDM-Cell.The experimental results show that the method can generate NuDM-Cell adaptively in different environments,and the localization accuracy can reach 95.7%,and the average localization time is 5.4ms.3)A multi-resolution hierarchical map multi-lane oriented construction and localization method is proposed and implemented.Firstly,a hierarchical annotation algorithm based on priori information and NuDM-Cell map is proposed to form a multi-resolution hierarchical map multi-lane oriented,and then a hierarchicallocalization network based on ResNet is proposed to realize lane localization by shallow features and intra-lane Cell localization by deep features.Experimental results show that this method can achieve better real-time localization effect,localization accuracy can reach 96.78%,and the average localization time is 7.3 ms.