Research On Binocular Vision Depth Recovery With Spatial Pyramid Vision Expansion Network

Today,the safe travel problem solved urgently in front of people is to design an intelligent driverless system and assisted driving system,which can reduce the traffic accident rates with human factors.The primary problem with the intelligent system is the perception of the road environment.Computer vision has a rich information,low power consumption,low cost and environmental sensing performance comparable to laser radar,which has become a research hotspot of environmental information perception.However,depth recovery technology of binocular vision still faces many difficult problems.ill-conditioned areas,such as occlusion area,repeated patterns area,no texture area and surface glare reflection area,are difficult to recover accurately depth information.In this regard,this paper proposes smart car binocular vision depth recovery algorithm with Spatial Pyramid Dilated Network(SPDNet).There are two main research aspects in this paper.1.Dark channel prior defuzzing algorithm by region segmentation optimizationThe first research aspect uses the dark channel prior defuzzing algorithm by region segmentation optimization to defuzzy and denoise the image acquired and corrected.Air pollution by road dust and car exhausting,and due to weather factor,a large number of particles are suspended in the air,images captured by camera are fuzzy and small contrast,which seriously affects the normal operation based on optical vision equipment.Defuzzing image based on the dark channel prior algorithm caused halo distortion in the sky region.To this end,this paper proposes an improved the defuzzing algorithm based on algorithm of regional segmentation optimization.Firstly,the fuzzy image is initially segmented by K-means algorithm,and then the local region of the atmospheric light value is selected as a mask,and the value corresponding to the mask region position of the K-means segmentation category is calculated as the category label value of the sky region.Sky area segmentation binary map is obtained.Furthermore,the sky region segmentation binary map is used as a mask,rough transmittance map is processed by regional optimization,which corrects the underestimation of the dark channel prior algorithm for the sky region transmittance map.Finally,the fuzzy image,the optimized transmittance map and the atmospheric light value are brought into the fuzzy image degradation model to restore the clear image,which lay the foundation for the next step of 3D depth recovery.2.Binocular visual depth recovery algorithm for unmanned vehicles with spatial pyramid dilated convolution networkThe second research aspect is a binocular vision depth recovery algorithm with SPDNet proposed in this paper,in order to recover depth information on the pre-processed left and right image pairs.Depth recovery for ill-posed areas,such as occlusion regions,lack of texture regions,sunlight reflection regions,etc.,Combining the context semantics of the global and local regions of the image,it facilitates accurate depth estimation.Because of the limited view fields of small-scale convolution kernels,it is difficult to merge global and local context semantics separately from the pixel level.Therefore,the SPDNet algorithm is proposed in this paper.Firstly,the residual feature convolution network,or ResNet50 is used to extract image features,then,the pyramid multi-level expansion convolution module(SPDModule)is proposed,which increases receptive fields of convolution kernel in order to merge context semantics of the global and local regions,and retains the location information of the regions.Then,the feature maps outputted by the ResNet50 network and the feature maps outputted by the SPDModule network are vertically superimposed.Finally,the feature maps outputted by the left and right image channels networks are also superimposed into 4D cost volumes in the vertical direction.The depth recovery is performed by the stacked hourglass 3D CNN network,which better solves the problems of difficult to recover accurately depth information from the ill-posed areas.The experimental results showed that the proposed algorithm eliminated the slab error depth estimation regions,compared with the average percentage of outliers for depth pixels of the Pyramid Stereo Matching Network algorithm,The ones by the algorithm proposed in this paper is reduced by 0.26%.The experimental results proved that the depth recovery algorithm proposed in this paper was better than the comparison algorithms,which showed the feasibility and effectiveness of the proposed algorithm.