AUTONOMOUS AERIAL MOTION PLANNING USING IMAGE PROCESSINGCOMPUTER VISION

Abstract

Autonomous aerial motion planning in a dynamic environment is a complex task, whereas recent advancement in the theory of dynamics, kinematics and computer vision demonstrated a remarkable potential for futuristic autonomous aerial platforms. Research is focus on generating real time navigations based on ordinary camera images. It takes run time images as an input and estimate depth from a predefined tag (Region of Interest) within the image; finally after position localization it plans a path to the goal point. Algorithm uses tag to update its orientation and to compute the distance. 2D images can be used to find the depth of a particular object in the image. Initially the shape of the marker is fixed and is pre-defined but as the research expands improvements will make the algorithm to work autonomously using everyday objects as tags. Feature detection and extraction are well known techniques in the computer vision and image processing. Already many algorithms are developed in this field. Most of these algorithms are limited to certain frame of reference. Such as corner points, lines, binary features, boundary traces extraction or detection. Likewise derived algorithm has its own novelty element. Algorithm independently not only classifies a pre-defined tag from a single 2D camera image but also calculates distance to that object with accuracy. Designed algorithm makes use of an easy to use graphical user interface developed to demonstrate results. Graphical user interface is designed in the Matlab. The research shows that single camera depth estimation can be achieved using polynomial curve fitting approach. For a said tag with known dimensions one can determine a fixed equation which can then be used to find any random distance. The approach is efficient and can effectively be applied to any indoor navigation or motion planning algorithm. Approach is not defendant upon expensive equipment and overall computational cost is also low. Ultimately the algorithm is capable of producing the output in just a few seconds with almost 91.84% accuracy.