Design And Research Based On Visual System For Blind Guiding Robot

Blind people face more difficulties and risks when they travel outside,which leads to their limited range of activities in a very small area.The research and development of blind guide robot will provide great convenience for blind people to travel outdoors.The visual system can detect objects and feedback them to blind people through human-computer interaction,which can make blind people get more information in complex and changeable outdoor environment,so as to avoid the risk that may occur.This paper aims at the target detection task of visual system,from five aspects:the hardware design of guide blind robot,the construction of visual system,the production of data set,the transplantation of neural network model and local path planning.The main contents of this paper are as follows:(1)The distance information and yaw information were collected by using RaspberryPi as the upper computer of the blind guiding robot.The visual system is built in the upper computer,and the object detection and recognition are carried out by using the image information captured by the visual sensor.The real-time communication between the upper computer and the lower computer is completed,and the data of multiple sensors is read.(2)By analyzing the task of object detection in visual system,ten object categories for object detection in outdoor environment are determined.According to the format of VOC2007 data set,the data sets of manhole and stoneballbarrier are made.The pictures are collected from the field.The data set includes 2147 pictures and 3006 target objects.(3)In order to run the target detection algorithm in RaspberryPi’s visual system,the endto-end neural network model YOLO v4 is used to train and transplant the model.Combined with the layer structure of YOLO v4 model,the residual module based on dilated convolution is used to improve the main feature extraction network,and the pooling operation is added after the residual module.By reducing the number of residual modules,the layer structure is optimized,and the training speed of neural network is improved.By reading the layer structure information in cfg file,the model file under Keras framework is transformed into the model file under DarkNet framework.(4)The artificial potential field method and dead reckoning are used to realize the local path planning of the blind guiding robot.Aiming at the situation that the artificial potential field method is trapped in the local minimum,the obstacle avoidance strategy of the blind guiding robot is proposed.The simulation experiments are carried out by using MATLAB and V-rep simulation environment to verify the obstacle avoidance effect in a variety of scenarios.(5)The training of neural network model is completed by using the established data set,and the evaluation of neural network model is carried out.The mAP value is 75.87%.Ten kinds of target objects are detected by using the pictures taken.At the same time,the real-time detection experiments of manhole and stoneballbarrier are carried out.According to the local path planning method,the obstacle avoidance experiment of the blind guiding robot is completed.