Vision-Based Autonomous Exploration And Scene Understanding For Micro Aerial Vehicles

In recent years, Micro Aerial Vehicle (MAV) has been from the original military product into personal consumption grade product. But most products are still dependent on the remote control to complete some simple missions. How to make the MAV have the ability of autonomous flight and complete more complicated and changeable task has become a hotspot in the field of MAV.In this paper, a new method of autonomous exploration and scene understanding is proposed. The small aircraft is equipped with computer, vision sensor, GPS system, so that it can obtain the current position and analyze the data in real time. Based the grid segmentation method, I use the SVM algorithm to identify the road area. Considering the need to make the path smooth and continuous, the system provides flight course information at a frequency of 3Hz. It is a great support to scene understanding.MAV scene understanding tasks include online estimation of geometric parameters, road blocking judgment, and the construction of outdoor scene map. Based on the color features, texture features of the data, with the SLIC super pixel segmentation algorithm to segment the data, in order to improve the accuracy of parameter estimation. With the calibration board designed by myself, search the relation among the aircraft flight altitude, ground object geometry and image resolution, to complete the geometric parameters estimation task. As for the consolidation on a wide range of outdoor scenes. We use the GPS system to record the flight path of the autonomous exploration, and make the real-time estimation of the road width, road length, record the obstacles in the scene information, and significant objects information. After that, the position of the above specific information will be marked on the flight curve and the construction of a large scale outdoor scene will be completed. A series of flight experiments have been carried out in natural outdoor scenes, the experimental results and data analysis show that the proposed method can meet the demand of Autonomous Exploration and Scene Understanding.