| Summary: | 碩士 === 國立臺灣大學 === 造船及海洋工程學研究所 === 89 === This paper presents a visual motion estimation system for biomimetic AUVs. To provide the function of object motion estimation through a machine vision system, after grabbed the sequences of video images, preceded processing is image improvement and segmentation. For image improvement, the Gaussian filtering is applied. Segmentation is when performed by using subtraction operations.
After the preceded processing, the scheme is composed of the optical flow analysis, movement calculation of moving objects camera calibration and stereo imaging. We choose two optical flow algorithms (Horn-Schunck method and SOAD) to calculate optical flow. For optimal movement calculation, the motion parameter algorithm is established, the optical flow is divided into object translation, rotation and distance change. Using a mathematical model for camera calibration, and combine with stereo imaging to obtaining the position and distance of the object in space.
To put programs to the proof, we make up the synthetic image pair. Use two optical flow methods to calculate the motion estimated. We considered different type and degree of movement on condition that the camera and background are still and only one object moving. To grabbing difference distance calibration image to calculate the calibration coefficients, and calculate the degree of accuracy of the coefficients. Last of all, calculate motion estimated for the real sequence images.
Results suggest that: (1) SOAD is batter than Horn-Schunck method. (2) The maximum of motion parameters error is 6%. (3) In the effective range, the error of distance estimate is under 3%. (4) The vision processing scheme which presented in this paper is feasible.
(keywords：AUVs, biomimetic, underwater robot vision, optical flow, camera calibration, stereo imaging, motion estimated)
|
|---|
