Experimental Comparison of Different Feature Detection Algorithms for UAV Obstacle Avoidance

For small and low-cost UAVs, real-time autonomous obstacle avoidance is a challenging problem. Due to the weight limit of carrying additional sensors such as radar, laser, and etc., vision-based autonomous obstacle avoidance has become a popular trend for small drones. In the past, several studies have used vision-based approaches for the three-dimensional reconstruction or depth estimation in obstacle avoidance. In this paper, the concepts of emulating human behavior to avoid an obstacle is used in the experiments. Two photos can be compared by detecting and matching the features with the principle of size expansion when an obstacle is approaching, and the matched features can construct a convex hull. This convex hull is treated as an obstacle which needs to be avoided. This method is fast and effective. In this study, several feature detection and matching algorithms are compared by experiments, and constructed convex hulls are used to compare the accuracy of obstacle detection. The larger the overlapped region between the convex hull and real object, the better the obstacle detection accuracy. This study analyzes and summarizes the experimental results to find which algorithm could obtain better accuracy and computation time, i.e. which could be more appropriate to utilized in UAV vision-based obstacle avoidance.