Design And Implementation Of A Computer Vision-Based Assistive Device For Blind People’s Mobility

Due to their visual impairment,blind people face great inconvenience in daily travel,and how to better assist them has become a social concern.Currently,there are various products available on the market aimed at assisting blind people with travel,including guide canes,handheld navigation devices,wearable navigati on devices,and guide robots.Although these products can provide some assistance,their feedback devices are mostly in the form of voice or vibration,and blind people can only passively follow prompts.Furthermore,due to immature technology and high costs,these products are not widely adopted in the market.To address this problem,this articl e proposes a computer vision-based device to assist blind people with travel.The device can segment the shape of the tactile pavement and convey the information to the blind person through servo motor rotation.Additionally,based on lightweight object detection algorithms and object tracking technology,obstacle avoidance and following functions are achieved.The main contributions of this article are as follows:(1)To address the problem of slow segmentation speed of traditional image processing methods,this article proposes the use of regions of interest to crop the raw input and reduce computation.To address the issue of poor denoising and shadow removal of traditonal filtering in the tacti le pavement segmentati on algorithm,a mean shift filtering method is introduced,and a shadow removal strategy is proposed based on the grayscale distribution histogram.The improved tactile pavement segmentation method has a fourfold speed increase,and the visual effects of segmentati on accuracy and shadow removal are significantly improved.(2)For the detection model,this arti cle prunes the neck feature fusion part of YOLOv8Nano,introduces PConv to the C2f module,and achieves a 1.5%and 3.4%speed increase、respectively,compared with the original YOLOv8-Nano model with an average precision loss of less than 1%.Furthermore,this article simplifies the tracking model and introduces ELAN and P-ELAN based on ELAN for comparative experiments.Compared with the ResNet18 feature extraction network,the training time is reduced by 30%and 33%,respectively,the model complexity is reduced by 74.5%and 82.8%,and the model parameter volume is reduced by 78.0%and 89.5%,with inference speed increasing by 81.2%and 96.2%.(3)Based on the Jetson Nano edge computing platform and Arduino microcontroller,this article designs a wearable device to assist blind people with travel.The device can determine the shape of the tactile pavement through dot matrix segmentati on and convey the information to the blind person through servo motor rotation.In addition,the device integrates object detection and tracking algorithm models to achieve tactile pavement reproduction,obstacle detection,and person following functions.