Monocular vision based obstacle detection

Detecting and preventing incidents with obstacles is a challenging problem. Most of the common obstacle detection techniques are currently sensor-based. Mobile robots like Small Unmanned Aerial Vehicles (UAVs) are not able to carry obstacle detection sensors such as radar; therefore, vision-based methods are considered, which can be divided into stereo and mono techniques. Mono methods are classified into two groups: Foreground-background separation, and brain-inspired methods. Brain-inspired methods are highly efficient in obstacle detection. A recent research in this field has focused on matching the Scale- Invariant Feature Transform (SIFT) points along with SIFT size-ratio factor and area-ratio of convex hulls in two consecutive frames to detect obstacles. However, this method is not able to distinguish between near and far obstacles nor the obstacles in a complex environment and, thus, is sensitive to wrong matched points. This paper aims to solve the aforementioned problems through using the distance-ratio of matched points. Then, every point is investigated for distinguishing between far and near obstacles. The results demonstrated the high efficiency of the proposed method in complex environments. The least achieved accuracy of the algorithm was 60.0%, and the overall accuracy was 79.0%.