Visually impaired people need assistance to navigate safely, especially in indoor environment. This research developed an obstacle avoidance system for visually impaired using Kinect’s depth camera as the main vision device. A new approach called auto-adaptive thresholding is proposed to detect and to calculate the distance of obstacle from the user. The proposed method divides equally a depth image into three areas. It finds the most optimal threshold value automatically (auto) and vary among each of those areas (adaptive). Based on that threshold value, the distance of the closest obstacle for each area is determined by average function. To respond the existence of the obstacle, the system gives sound and voice feedback to the user through an earphone. The experimental result shows that execution time and error of the system in calculating the distance of the obstacle are 12.24 ms and 130.796 mm respectively. Evaluation with blind-folded persons indicates that the system could successfully guide them to avoid obstacles in real-time condition.