電腦視覺為基礎之多部位人體追蹤系統設計

Abstract

本研究我們提出一個在錄影序列中之多部位的人體追蹤系統，首先我們利用一個背景模型偵測並切割出錄影資料中之人體，由於背景影像通常成區塊變化，空間特性可用來表示背景的外觀，為了要建立背景外觀的空間特性的模型，我們為成對的點上之組合顏色建立混合高斯分布的模型。當人體被偵測及切割出來後，接著我們使用人體部位外觀當作特徵並且使用粒子濾波器作為核心來追蹤此人體，我們採等化之顏色統計表當作粒子濾波器中使用的外觀特徵以強化不同物體的鑑別率，為了建立可穩健區別目標物及背景物體的追蹤器，我們同時使用目標的模型和背景模型來計算目標物的相似度，為了對抗背景及目標物的外觀變化，背景模型及目標物模型都是可適應變化的。在一個粒子濾波器中，當粒子數量很多時，特徵抽取的過程會有多餘的重複計算而沒有效率，為了加速特徵抽取，我們為每張影像建立了數張累加的統計圖，每一個粒子的顏色統計表可因此在常數時間中被計算出來。當追蹤人體的時候，我們會把人體切割為三個部位：頭、軀幹、臀腳，這三個部位會分別被表示成為內縮的矩形並用粒子濾波器來追蹤，因為這樣的處理，我們可以檢查這三個部位的一致性以減少可能的追蹤失敗，當追蹤過程中追蹤狀態被更新後，我們會用支持向量機（SVM）來偵測追蹤錯誤並且判斷不正常的部位，假如只有一個部位不正常，我們會校正這個不正常的部位並利用系統動態模型來追蹤此部位，假如兩到三個部位不正常，我們就會從出此三個部位的預估位置重新初始化這三個部位的追蹤。實驗結果顯示，我們提出的背景模型可以有效的偵測背景有改變時及物體在原地移動時的移動物體區域，跟高斯背景模型和混合高斯背景模型比較，我們提出的方法可以抽取出更完整的移動物體區域。人體追蹤的實驗顯示出，我們提出的三部位人體追蹤及錯誤校正可以正確持續追蹤95%的人高達105個畫面，考慮人體部位追蹤，我們提出的系統可以持續追蹤頭部、軀幹、臀腳這三個部位105個畫面的正確率分別高達95%、83%、91%，和整個人體視為一個部位的追蹤作比較，正確率提升20%，這個結果顯示出這個系統是一個有效的追蹤系統。

The study presents a multi-part human tracking system in video sequences. First, we detect and extract humans in a video according to a background model. Since background images usually change in blobs, spatial relations are used to represent background appearances. To model the spatial relations of background appearances, the joint colors of each pixel-pair are modeled as a mixture of Gaussian (MoG) distributions. After the human is detected and extracted, we then track body parts of the human by using appearances of these parts as the features and using particle filters as the tracking kernel. In the particle filter, we adopt color histograms as the appearance features and use a specific histogram mapping to enhance the discriminability between different objects. To form a robust tracker that can distinguish target objects from background objects that have color distribution similar to those of target objects, we calculate the target similarity from both the target object model and the background model. To handle the appearance variations of background and target objects, both the models of the background scene and the target object are adaptable. In a particle filter, when the number of particles is large, the feature extraction is repeated redundantly and inefficiently. To speed up feature extraction, we create a cumulative histogram map from each image. The color histograms of each particle can then be extracted in constant time. When tracking a human, we decompose the human body into three parts: head, torso, and hip-leg, represent them by three shrunk rectangles, and track them by particle filters. In this way we can reduce possible tracking failures by checking the consistency of states among these three parts. After the tracking states are updated, we use support vector machines (SVM) to detect tracking failures and abnormal body parts. If a single part is abnormal, we adjust its position and use the system dynamic model to track the abnormal one. If two or three parts are abnormal, we re-initialize the tracking process of the three parts around their predicted positions. Experimental results show that the proposed background model can be used to efficiently detect the moving object regions when the background scene changes or the object moves around a region. By comparing with the Gaussian background model and the MoG-based model, the proposed method can extract object regions more completely. The experimental results of human tracking showed that the proposed three-part tracking system with failure detection and correction can track correctly about 95% persons until the 105th frame. With respect to the body parts, our system has about 95%, 83%, and 91% tracking rates for the head, torso, and hip-leg parts respectively until the 105th frame. The tracking rate of a human increases 20% comparing with that of the whole-body tracker. These rates show the effectiveness of the proposed system.