Illumination Invariance In Road Scenes And Its Applications

With the development of economy and urban traffic,driving conditions have become more complicated,causing a large number of traffic accidents.As a result,researches on unmanned vehicles and Advanced Driver Assistance Systems become popular.This also promotes the development of Pattern Recognition,Computer Vision,Intelligent Control and etc.A tough and critical problem remains in such systems that how to use images to achieve an accurate environment sensing in complicated urban areas,as the presence of shadows in road scenes is a tricky issue to deal with.To help solve this problem,we focus on the detection and removal of shadows and propose two solutions in this paper.Below are our main contributions:(1)Based on the model of image formation,we propose an illumination-invariant feature to eliminate the effects of shadow in road scenes.As the chromaticity of a ho-mogeneous dielectric distributes as a planar structure,we use Principal Components Analysis(PCA)to extract the structure and define the illumination-invariant feature.Then,this illumination-invariant representation is applied to two road applications:the vanishing-point detection and road segmentation.The improvement of performance in both two applications verifies the validity of our illumination-invariant feature.(2)We propose a fully-convolutional deep network,U-shadowNet,which can both accurately detect and remove shadows in general scenes in an end-to-end way.We enable the network to extract higher-level features of the images and improve the performance of the network by introducing residual modules.Several skip-connections are built in the network.This architecture helps U-shadowNet combine the shallow and deep features and grab the global semantic information of the image7 which is crucial to both the detection and removal of shadows.(3)Based on the shadow detection framework of U-shadowNet,we perform shadow detection on the complicated road scene.We then exploit Conditional Random Fields(CRFs)to optimize the preliminary shadow detection results.So that we can obtain the shadow matte based on the shadow edge and finally eliminate the shadows on road.