Design Of A Visual Assistant System For Blind People Based On Deep Learning

At present,there are 253 million visually impaired people in the world,and their lives and travel have been greatly troubled by visual obstacles.Traditional auxiliary blinding tools are unable to identify dynamic vehicles,which leads to the limited effect of auxiliary blindness.In China,some vehicles,including electric bikes and shared vehicles,are even driving on the blind track,which poses a great threat to the safety of the blind.Therefore,the research of non-contact blind people's visual aid with obj ect recognition and distance estimation has important social and engineering significance.Most of the existing solutions rely on lasers,ultrasonic and other devices to detect obstacles.They cannot simultaneously measure multiple obstacles and identify obstacle categories.Therefore,it has great limitations.In this thesis,we take advantage of target detection and recognition in the computer vision system to realize the simultaneous detection of multiple targets without adding additional sensors,which brings convenience to the travel of the blind.Specifically,this thesis uses Client/Server(C/S)architecture to realize an intelligent wearable auxiliary blind hardware device.The client is a simple hat that takes pictures of the scene in front of blind person through the camera on the hat and transfers data to the server through raspberry pi..On the server side,after comparing the actual detection results of SSD and YOLOv2 algorithm,YOLOv2 algorithm is used to detect and identify the obstacles in front of the blind.In allusion to the shortage of YOLOv2 cannot identify high risk electric vehic les,the pictures of various types and angles of electric bicycles were collected to retrain the corresponding network model parameters,and 86.5%accuracy rate was achieved for identifying the electric bicycles.Besides,this thesis proposes a new method for quickly estimating the distance of multi-moving targets using monocular cameras.Compared with other ranging methods used by other auxiliary blind systems,the proposed distance measuring method features advantages of low computational complexity,high speed,and no need for additional distance measurement.Finally,the system performance is tested.The testing results show that the proposed intelligent wearable device can effectively make the blind people "see" the scene in real time,and can effectively identify and estimate distance of five kinds of high risk objects including cars,bicycles,motorcycles,electric bicycles and pedestrians.