A New Real Time Environment Perception Method Based on Visual Image for Micro UAS Flight Control

Recently, it has been a popular issue that using visual perception for autonomous navigation in Unmanned Aerial Vehicle (UAV). However, in the field of computer vision, the restoration of the three-dimensional shape from the two-dimensional image video stream captured by camera is the central issue, which includes the feature point selection, tracking studies, and structural problems of uncalibrated image sequences, motion inference structure technology as well as Structure from Motion technology (SMF). In theory, a variety of innovative features pixel selection method and tracking methods have been excavated, for multi-angle geometry, projective reconstruction, the fundamental matrix camera self-calibration technique are estimated to have a considerable development. In application, with the full demonstration of the practicability of UAV, various techniques emerge one after another, aiming at autonomous navigation. Many of the applicable fields have demanding requirements to three-dimensional graphic data. How to obtain the exact three-dimensional environmental information timely have become the current topics. The fact is that in actual flights, the flying area of small UAV is comparatively very complex. For insurance of normality, we have to make the small UAV compute quickly so as to produce quick reaction to the unknown terrain. However, the current small UAV are generally with a small computing capacity and a low computing ability. To assure the normal flights of small UAV, currently we have to lessen the navigation accuracy for exchange of the quick computation. The thesis firstly analyzes the existing advantages and disadvantages of the small UAV. Based on the current techniques, the approximal tracking algorithm has been extracted, which is the most appropriate to the current need. The remaining topics will be discussed in future studies. Firstly, SUSAN corner detection algorithm is adopted to detect the sequence of two-dimensional image so as to obtain special points, in the arbitrary image frame, which enjoy rotation invariance and do not vary with the change of light. Also the methods to improve SUSAN algorithm are proposed. Next, KLT feature tracking algorithm is used for the feature corner tracking of adjacent image frames, and the feature sequence is thus formed. Finally, the method introduced in this thesis is verified valid through simulations of practical video streams.