摘要：AbstractThe increasing application of unmanned aerial vehicles (UAVs) in unstructured, natural environments raises the demand for robust obstacle avoidance systems that work in realtime. In view of this problem, we present a control algorithm for quadrotors based on monocular vision that is specifically designed for obstacle avoidance in forest environments. The algorithm presented is an enhancement of an existing algorithm, originally proposed and tested for usage with rovers, that uses a weighted combination of texture features to compute distances to nearest obstacle in various longitudinal strips of the image frame. The weights are pre-computed by means of supervised learning of correspondences of the features to ground-truth distances processed on frames derived from a simulated forest environment. The modifications proposed on the original algorithm show significant improvement in its obstacle-distance estimation accuracy and computational efficiency as observed from actual autonomous flying experiments carried out on an off-the-shelf quadrotor. The modified algorithm works at 15 frames/sec for a frame size of 160 x 120 pixels as profiled on an Odroid XU4 mini-computer. Results from both simulated images and real videos have been presented.
