In this paper we propose a navigation assistant for visually impaired people, which uses computer vision techniques and is integrated on a wearable device. The system makes it possible to detect and recognize, in real-time, both static and dynamic objects existent in outdoor urban scenes without any a priori knowledge about the obstruction type or location. The detection system is based on relevant interest point extraction and tracking, background/camera motion estimation and foreground object identification through motion vectors clustering. The classification method receives as input image patches extracted by the detection module, performs global image representation using binary VLAD and prediction based on SVM. The feedback of our system is transmitted to visually impaired users through bone-conduction headphones as a set of audio warning messages. The entire system is fully integrated on a regular smartphone. The experimental evaluation performed on a set of 20 videos acquired with the help of VI users, demonstrates the pertinence of the proposed methodology