A Novel Trail Detection and Scene Understanding Framework for a Quadrotor UAV With Monocular Vision

Vision-based trail detection and autonomous scene understanding play a key role for unmanned aerial vehicles (UAVs) working in complex outdoor environments such as isolated disaster sites. This paper investigates the problems associated with trail detection and tracking, as well as autonomous scene understanding using a quadrotor UAV. A framework that integrates support vector machine-based trail detection with a trail tracker is proposed to accomplish trail direction estimation and tracking at a low cost of computation and in real time. To accurately perform online parameter estimation, a performance test is designed and implemented to evaluate the accuracy. Moreover, the simple linear iterative clustering superpixel segmentation algorithm is utilized in the proposed system framework to guarantee the scene segmentation accuracy. Visual detection for significant objects or people is implemented by using single shot multibox detector algorithm. A series of experiments are conducted by using a quadrotor platform DJI M100 and experimental results show the validity and practicality of the proposed approach.