多攝影機資訊整合之人員定位與誤差分析

Abstract

在本論文中，我們實作一套基於電腦視覺的定位系統，該系統運用到平面投影轉換關係和三維幾何關係，達成對影像中的目標物進行定位。然而實際定位的過程中仍將存在著某些誤差，例如徑向畸變誤差、人為測量誤差和取像誤差，這些誤差將會使得定位的準確度下降。因此我們對於定位過程中可能產生的誤差加以分析與改善，減少這些誤差對於定位的影響，我們所採取的方法包含對影像作校正和微調測量的位置。另一方面，我們可以利用分析的結果進一步地由多攝影機中挑選出，具有較低定位誤差之攝影機配對，以獲得較佳的定位結果。由虛擬場景和實際場景的實驗結果可看出，利用誤差分析的結果能使我們成功地整合多攝影機資訊，並取得較為準確、穩定之定位結果。

In this thesis, we implement a vision-based localization system that can locate objects in the scene from their images by homography and 3D geometry. In reality, the object localization may not be always accurate due to errors arising from radial distortion of cameras, noises in the imaging process, and errors associated with manually measurement, etc. We analyze these errors and try various ways, including simple camera calibration and fine-tuning the measurement data, to reduce the degradation in localization accuracy. On the other hand, according to these analytic results, the system can suggest an appropriate pair of cameras in a multi-camera environment to achieve more accurate localization. Both synthetic and real scene data are used to verify the implemented localization system. Experimental results show that the proposed approach can indeed reduce the localization error and improve system stability.