利用多重亮度紅外線打光器之夜間人物偵測

Abstract

一般夜間監控常因光源不足，無法看清較遠之前景物，而在光源充足的情況下，也可能因人物離攝影機過近，造成影像過曝而無法辨識其面貌。對此，我們使用週期性改變光源強度之多重亮度紅外線打光器，輔助夜視攝影，以得到多種亮度影像，並對此新式夜視影像做分析，以期同時達到遠景物偵測正確與近景物清晰辨識之目的。針對多重亮度打光器所拍攝之新式影片，本論文提出三種不同前景區域偵測方法，其中前兩種方法以影像亮度分群法為基礎，自動找出影像光源亮度變化週期，與影像亮度群組，進而切出前景區域。第三種方法提出週期性最大最小值模型(periodic min-max model)，在低計算複雜度下，不需經過亮度分群，可做到快速前景區域偵測。實驗結果顯示，第一及第三種前景區域偵測方法，在各種測試影片中，平均正確率達到90%以上，且能得到正確偵測遠景人物，又能清楚顯示過近人臉之效果。

In nighttime video surveillance, far objects are often hard to be identified due to poor illumination conditions while near objects may be whitened due to over-exposure. Therefore, we used a multiple intensity IR-illuminator that provides multiple illumination levels periodically as a supportive light source. By using the IR-illuminator, images with different degrees of exposure can be obtained. Accordingly, we can detect far human correctly and display near face clearly in the meantime, which is better than using fixed illumination. We analyze multiple intensity videos and propose three foreground detection methods. The first two methods can extract foreground regions by illumination clustering, which can find the period of illumination variation automatically. The third method, based on a periodic min-max model, has lower computational complexity and real-time performance and does not require illumination clustering. Experimental results show that the first and the third methods can achieve more than 90% of average accuracy in foreground object detection for various test videos.