Research And Implementation Of Image-based Blind Channel Semantic Segmentation Technology

According to the data of the second survey of disabled people in China,the number of visually impaired people in China exceeds 12 million,which is the largest in the world.In order to ensure the travel of visually impaired people,many cities have set up blind sidewalk areas on the sidewalks,which can provide visually impaired people with tips for moving,turning and arriving.In recent years,the emergence of electronic blind guide systems based on image recognition has brought more convenience to the visually impaired.Blind sidewalks segmentation is an important part of the electronic guide systems.In order to improve the efficiency and quality of blind sidewalks segmentation,this paper mainly does the following research:(1)This paper proposes an improved blind sidewalks segmentation algorithm based on traditional digital image processing technology.In the proposed improved algorithm,firstly,the features extracted from the color histogram are pre-classified by SVM,and the blind sidewalks pictures are divided into two types: color blind sidewalks and texture blind sidewalks.Secondly,the color blind sidewalks converted into the HSV color space are segmented by improved OTSU multi-parameter fusion;the texture blind sidewalks are modified by Gaussian difference preprocessing,the gray level co-occurrence matrix extracts the features,and finally the K-means clustering segmentation is used.Because of the accurate classification and segmentation algorithm for blind sidewalks,the segmentation effect under the influence of light and shadow is greatly improved.(2)In order to make the blind sidewalks segmentation algorithm more adaptable to the scene,this paper proposes an improved network model M3-Unet,which combines the characteristics of convolutional neural network U-Net and Mobilenet.In order to obtain the training effect in small data sets and reduce the consumption of running resources,the improved model uses the Mobilenet network with strong pattern recognition ability in the traditional U-Net coding layer to perform feature extraction.The experimental results show that using the M3-Unet model can extract more semantic information than the existing deep learning algorithms,which can effectively improve the blind sidewalks image segmentation effect and adaptation range.(3)In order to realize the application of the improved algorithm on the embedded platform,this paper builds an embedded platform based on NVIDIA Jetson nano,and transplants the proposed M3-Unet improved model on the platform to achieve real-time blind sidewalks segmentation.