Blind Navigation using Deep Learning-Based Obstacle Detection

Abstract

Blind navigation has become a challenging task in the present. Blind people cannot detect and avoid obstacles similar to the people with good vision and they need guidance to avoid such obstacles. The white cane is the most widely used device by many blind navigators to detect and avoid obstacles. But with the limited reachability of the white cane, it is not possible to detect all the potential threats to the navigator. Therefore, the white cane is not an adequate aid to navigate safely. To secure the safe and independent navigation of the blind people, more insights of their current surroundings must be provided. This study proposes a novel approach for obstacle detection based on deep learning to assist in blind navigation. In this study, a prototype was developed using deep neural networks (DNN) for obstacle detection and distance estimation due to real-time performance and high accuracy of DNNs. To train the DNN for obstacle detection data was gathered using a simulation environment. The output of the obstacle detection model was used to estimate the distance of the obstacles. The final result by combining the feedback of obstacle detection and distance estimation is communicated to the user via audio queues. The prototype system is deployed in a smartphone and the real-time video stream captured through the smartphone camera is processed to detect obstacles. To train the DNN for obstacle detection SSD MobileNet Architecture was used and the data to train the DNN was generated using a simulation environment. To estimate the distance of the detected obstacles, DNN based MonoDepth algorithm was used. The mean average precision (mAP) value of all the classes of the DNN for obstacle detection reached more than 70%. Higher accuracy and a higher speed for obstacle detection can be achieved by the system prototype but there is a latency when estimating distance. The usability and the effectiveness of the prototype system exceeded 65% according to the feedback from the usability evaluation.