基於異質攝影機之自動車夜間視覺定位與誤差分析

Abstract

這篇論文的主要目的是建構一個整合異質攝影機之移動型夜間監控系統。此系統包含近紅外線攝影機、中紅外線攝影機、自走車，並利用布設在場景中的近紅外線LED參考點，依使用者需求移動並進行即時的目標物（人物或動物）定位。系統先由從中紅外線影像中擷取具較高溫度的目標物位置，並從近紅外線影像中擷取近紅外線LED參考點位置。接著根據單應性轉換與平面座標轉換等原理，由影像中目標物的位置計算出其位於場景中的位置。為了增加定位之精確度，我們亦探討了上述影像擷取與座標轉換過程間可能發生的誤差，以及可能的改善方式。實驗顯示藉助此種誤差分析，本論文所提出的監控系統可以完成更為準確、穩定之目標物定位。

The main purpose of this thesis is to construct a mobile system that integrates heterogeneous cameras, a near-infrared and a mid-infrared camera, for nighttime surveillance. The system, which consists of these cameras and a mobile robot, can locate target objects, e.g., human or animal, using some reference LEDs in the scene. The system first detects high temperature objects and the reference LEDs in the mid-infrared image and near-infrared image, respectively. The localization of target object can then be performed efficiently by homographic transform and planar coordinate transformation. In order to reduce the localization errors, we also study how errors may be generated during the localization process, and possible ways to reduce their influences. Experimental results show that, with the help of such analysis, more accurate and reliable object localization can be achieved with the proposed surveillance system.