Moving Objects Detection by Image Methods on a Moving Platform
| Autor: | Chen, Cheng-Che, 陳正哲 |
|---|---|
| Rok vydání: | 2012 |
| Druh dokumentu: | 學位論文 ; thesis |
| Popis: | 100 Due to the change of pixels from each captured scene with time, moving objects cannot be detected by the traditional optical flow or frames difference methods. Two methods of moving objects detection on a moving platform are proposed in this study. One is “Static scene reconstruction”, a generalized method, and the other is “Single camera feature scale differentiation” for detecting approaching moving objects from backward camera. Both methods are tested by the image set captured from the real outdoor scene. Static scene reconstruction is verified first by the indoor scene with plentiful and distinct feature points which enhance the robustness of feature point matching. Static scene transform matrix can be estimated by the sequential feature match. Feature points on moving objects are defined as feature points with large distance between the positions of feature point predicted by the static scene transform matrix and the corresponding detected feature points. Some modifications are needed when applying Static scene reconstruction for the outdoor scene. A normalized criterion is proposed to enhance the matching between the ambiguous features from the outdoor scene. In addition, the distant difference between the objects is larger in the outdoor scene than that in the indoor one. The shape of the static scene can be improperly changed by the unconstrained scene transform matrix accordingly. A method of computing the rotation matrix by far feature points in an image and the translation vector by near feature points is proposed. The process by this proposed method is faster than that by the rigid body transformation method which estimates the transform matrix by mean square error. Moving objects are successfully detected in the outdoor tests. For objects that approach the moving platform backwards, a method of feature scale differentiation by single camera is proposed. Feature points of scale increasing are selected as moving object candidates based on this method. Moving objects can therefore be found by refining the moving object candidates after the consideration of the spatial relationship with the adjacent features. Approaching moving objects are successfully detected in the outdoor scene, and the processing frequency is 4.4 fps, which is faster than the acquiring frequency of 2 fps. |
| Databáze: | Networked Digital Library of Theses & Dissertations |
| Externí odkaz: |
|