Research On Fine Positioning System For Blind Sidewalks Based On Binocular Stereo Vision

Vision plays a very important role in our daily lives.However,with more than 2.2 billion people blind or visually impaired worldwide,and with the rising number,the visually impaired population is gradually increasing.In their daily travels,visually impaired people are unable to avoid obstacles and discover walking paths through visual perception of their surroundings,so they need blind sidewalks to guide their way forward.Therefore,it is very important to help visually impaired people accurately identify the location of the blind sidewalk.In the area of blindness-assisted localization,most of the existing assistive technologies aim to avoid obstacles or plan paths to reach specified locations,without giving full play to the role of blind sidewalks.Hence,a fine positioning system for blind sidewalks based on binocular stereo vision is proposed in this paper to improve the safety of visually impaired people’s travel,as follows:A ME-UNet network for accurate segmentation of blind sidewalks from complex backgrounds is proposed to address the problems of inaccuracy and few segmentation types in the existing blind sidewalk segmentation methods.Firstly,a diverse dataset of blind sidewalk semantic segmentation is constructed to increase the adaptability of the blind sidewalk segmentation network for different types of blind sidewalks.The dataset includes four types of blind sidewalk images: yellow blind sidewalks,gray blind sidewalks,red blind sidewalks and metal color blind sidewalks.Then,the correlation between feature channels is enhanced by adding the SE attention module(Squeeze-and-Excitation,SE)in the shallow part of the network.Next,the Multi-scale Feature Capture Module(MFCM)is integrated to extract the feature information of different perceptive fields.Finally,multiple Enhancement of Dependency Attention Modules(EDAMs)are adopted in the network to enhance the contextual information fusion.A fine blind sidewalk localization algorithm based on binocular stereo vision is proposed based on the accurate segmentation of the blind sidewalk area.The blind sidewalk area is segmented by the ME-UNet model,edge detection is performed on the area,and the edge information is obtained by refining the edges.The spatial coordinates of key points are calculated by the blind sidewalk key point determination algorithm,and the distance and angle information between the blind sidewalk and the visually impaired person are finally acquired,which could produce more accurate travel assistance hints for the visually impaired person.Experimental results show that the proposed ME-UNet network achieves 94.24% and 97%in segmentation accuracy MIo U and MPA,which are 2.37% and 0.83% higher than those of the original U-Net respectively and can segment blind sidewalks more precisely.In the proposed fine blind sidewalk positioning system,the relative error of the measured distance is within ±2% and the absolute value of the angle error is less than 4°,which can meet the requirements of practical applications.