Obstacle Detection, Depth Estimation And Warning System For Visually Impaired People

The work aims to develop a system which assists the blind people in sensing and identifying obstacles on their way. It also estimates the distance towards them. We use Yolov2 to train the system. With this, an image description is produced along with its distance from the person. The system would then present the person with an appropriate audio response. The range information is obtained by merging the output of object detector with that of an ultrasonic sensor mounted on an Auridino Uno. We evaluated the system both indoors as well as outdoors. The system can detect the object using the pre-trained deep neural network with 97% accuracy. We want to extend this work into a portable device like mobile phone by transfer learning of the model. The proposed system can identify more object categories along with depth estimation, with reasonable accuracy, while others can identify only the presence of an obstacle.