Remote Sensing Image Road Extraction And Road Network Constructio

With the rapid development of artificial intelligence technology,automatic driving technology has made great progress.Automatic driving technology relies on sensors’ perception of the surrounding environment to make decisions and control strategies.With the increasing complexity of automatic driving demand and scene,the demand for road network map is also increasing.Accurate road map can provide stable road structure information for the autopilot system,and make up for the lack of sensor perception due to other interference or distance constraints in environmental perception.At present,the construction of road network map mostly depends on the data collected by handheld or on-board GPS equipment,and then combined with manual knowledge to make the road network map.Although the road map built by this way is more accurate,it is impossible to obtain the prior road information of the target area in advance when the target area cannot be entered in advance.The road information of the target area can be obtained automatically by road extraction from remote sensing images,and the road network map can be constructed efficiently without entering the target area for data acquisition based on the results of road extraction.At present,road extraction from remote sensing images is mostly completed by semantic segmentation.The task of road extraction from remote sensing images faces many challenges: the road background is complex,and the road pixels have high similarity with some background pixels,so the road boundary is always unclear;The road usually presents a narrow and slender structure,occupies few pixels,but tend to span in the whole image,so the number of road pixels and background pixels is extremely imbalanced;Remote sensing images are imaged from a top view.Buildings,trees and their shadows near the road will occlude the road,thus damaging the connectivity and completeness of the road.In view of the above directions,this paper improves the existing road extraction algorithms of remote sensing images.In addition,in order to build the road network map of the target area without data acquisition in advance,a system is designed.The main contributions are as follows:(1)We design a High-resolution Feature Fusion Network(HRF-LinkNet)based on the classical model D-LinkNet of road extraction from remote sensing images,while retaining the intermediate module of D-LinkNet for extracting multi-scale features,layer by layer down sampling connection is added to the model,and more precise spatial features are introduced into deeper high-level semantic features to alleviate the problem of unclear boundary of road extraction results.In addition,due to the weak reusability of feature maps,up sampling connection is introduced to make the middle scale features participate in the decoding process of the model,so as to make full use of the reusability of the model.(2)To solve the problem that HRF-LinkNet still does not make full use of all scale features,we design the All-scale Feature Fusion Network(AF-Net)by adding the All-scale Feature Fusion Module(AF-Module)to the classical semantic segmentation model LinkNet to incrementally integrate the feature maps of all scales into the feature maps of each scale,so that the feature fusion results of each scale contain more accurate spatial information and richer high-level semantic information.In order to solve the problem of introducing unnecessary noise when fusing the features of all scales,we introduce attention mechanism into AF-module to enhance beneficial information and suppress useless information.In addition,in view of the extremely unbalanced of positive and negative samples in the road extraction task of remote sensing images,a composite loss function composed of weighted binary cross entropy loss and dice loss is constructed to help the model better achieve the road extraction from remote sensing images.(3)In order to convert the road extraction results of remote sensing images into the road map that can be used by the automatic driving platform,this paper designs a solution to convert the road extraction results of remote sensing images into the road map with format of Open Street Map,a high precision road map currently used by the automatic driving system,so as to realize the automatic road map construction of remote sensing images.Finally,it achieves the goal of efficiently completing the road network map construction without entering the target area of data collection in advance.