Embedded Blind Visual Aid System Based On Jetson Nano

With the rapid development of computer vision technology and the progress of embedded GPU equipment,plenty of convenience had been brought to people.In the domain such as robot,security,automatic driving has made remarkable achievements,at the same time,the emergence of new equipment,new technology can bring more convenience to the group of blind.In recent years,many domestic and foreign scientific and technological workers devoted themselves to the research of blind auxiliary devices,aiming to provide assistance and support for the safety of blind people in travel and obstacle avoidance,and so on.However,they have many problems such as lack of convenience,high price,insufficient environmental cognition ability,and those devices are more suitable for outdoor environment.Most of the activity time and space of human beings are still in the indoor environment,especially for the blind group.How to effectively help the blind group to find the target they are interested in and improve their perception and cognitive ability to the surrounding environment has become an important topic,which is also the main content of this paper.Based on the above background,this paper researches and explores relevant technologies,and proposes an embedded visual assistance system for the blind based on Jetson Nano.The system uses target recognition algorithm and binocular ranging algorithm to provide visual assistance for the blind to fetch objects in the indoor environment and help them keep away from dangerous objects.The main work contents of this paper include three aspects: firstly,learning and researching the FASTER-RCNN algorithm,SSD algorithm and YOLO algorithm,and finally determining the YOLOV3 Tiny algorithm as the target detection algorithm adopted in this paper;secondly,improving the YOLOV3 Tiny algorithm,The accuracy of the model is improved by deepening the Tiny network layer of YOLOv3 and adding the attention module.Meanwhile,the Tensor RT inference acceleration framework is adopted to improve the speed of the model.Thirdly,the detection box coordinate information provided by the target recognition algorithm is combined with the depth map generated by the BM algorithm to obtain the depth information of the identified target,and the azimuth information of the identified target is obtained based on the detection box coordinate information.