影像式人潮偵測與估計之探討

Abstract

在智慧型監控系統中，人潮的分佈區域、時段、數量、趨勢等資訊可以提供許多不同層面之應用，如商場中之動線規劃及運輸系統中人潮數量偵測等。有別於傳統的紅外線感應器，本研究利用影像辨識技術取得人員數量的相關資訊，並結合相關的智慧型影像分析技術，可以提供一先進的影像監控服務。本篇論文提出一個高角度攝影機拍攝場景下之人潮偵測研究，我們利用基於高斯混合模型的背景分割方法，劃分出可能之人員分佈區塊，然後再利用各區塊之特徵來估計出場景中的人員數量。關於特徵的選用，首先我們先利用前景物之輪廓特徵來偵測人群之位置，接著再利用頭部顏色與形狀特徵來進一步偵測輪廓內部之人員，最後我們利用每個影格的前後關連性來排除錯誤偵測以及增加系統之正確性。經由多個真實場景之實驗統計，本論文所提出之人潮偵測方法正確率可達80%以上。

The detection of a crowd of people, and the analysis of its size as well as its distribution in space/time, has many applications in intelligent surveillance. For example, planning a path of shopping in a store or the detection in a transportation system relays on the above analysis. In contrast to using traditional infrared sensors, more interesting surveillance services can be provided through advanced image recognition technologies for people counting and location identification. In this thesis, we propose a crowd detection method for viewing a scene. We first segment the foreground region of a crowd with Gaussian mixture models. Subsequently, color and location attributes of head region are used together in people counting. Finally, the foreground area, as well as interframe relation of head regions, utilized to improve the robustness of the crowd detection, Experiments with real-world videos show that over 80% accuracy rate can be achieved with the proposed approach.