結合時間與空間資訊之場景變化偵測新簡易方法

Abstract

本論文的目標是場景變化偵測，其中包含了許多不同環境中的複雜問題如下 (a) Bad weather (下雨、下雪、大太陽、濃霧…等等)。 (b) Dynamic background (樹的搖晃、水面波紋、灑水噴泉)。 (c) Time of day (破曉、中午、黃昏、夜晚)。 (d) 室內或室外。 (e) Foreground aperture：內容均勻的物體移動時，其內部的Pixels無法被偵測，因此整個物體可能不被認為是前景。 (f) Sleeping or waking person：當一個人突然的停下靜止不動，就可能會誤判成背景；或是一個人從靜止狀態轉變成為移動狀態，則原本靜止的位置新露出來的真正背景就可能產生誤判。 (g) 不同的frame rates可能會有不同的學習速率。 (h) 背景是雜亂的，並非均勻的分佈。 首先，回顧七種現有的不同場景變化偵測方法，且從網站下載不同的Dataset當作我們測試的數據，然後我們定義出6種Pixels Type 如下 (1) Stationary pixel, (2) Quasi-Stationary pixel, (3) Cyclic pixel, (4) Quasi-Cyclic pixel, (5) Bimodal pixel, (6) Heterogeneous pixel. 我們設計了演算法來偵測這6種不同的Pixels Type，且用來做Background Model訓練。 在Background Model訓練完之後，若有新的Frames進來，則我們可以更新Background Model，因此我們可以處理slow lighting change問題。若是光線突然變化，則系統可以重啟，再重新訓練Background model，或是允許較大的強度偏移量來改善。 在電腦實驗中，我們使用了足夠的Dataset且準備實驗室來測試我們新的場景變化偵測方法，我們得到了中間結果與最終結果，且為了測試性能，我們也計算了Precision、Recall和F-Measure做為指標，整體來說，我們的方法表現良好，並能處理一些現有方法所無法處理的問題，像是Sleeping or waking person；此外，我們的方法可以使用較少的Frames數目做訓練，以快速適應場景變化。

The thesis addresses the issues of the scene change detection in various complex environments including (a) Bad weather (raining, snowing, sunny, and foggy, etc.), (b) Dynamic background (weaving tree, wavy lake/sea, fountain sprinkle), (c) Time of day (dawn, noon, sunset, night), (d) Locations in indoor or outdoor, (e) Foreground aperture where a homogeneously colored object moves, change in the (object) interior pixels cannot be detected. Thus, the entire object may not appear as foreground, (f) Sleeping or waking person when a person becomes motionless, causing a misclassification as background or a person motionless starts to move, causing both it and the newly revealed parts of the background appear to change, (g) Various frame rates causing different background learning rates required. (h) Cluttered background with non-uniform intensity distributions. To begin with, a review of seven classes of existing scene change detection techniques is conducted and various datasets of different nature are downloaded from the websites are selected as the test bed. We then define six types of pixels in a general scene, namely, (1) Stationary pixel, (2) Quasi-Stationary pixel, (3) Cyclic pixel, (4) Quasi-Cyclic pixel, (5) Bimodal pixel, (6) Heterogeneous pixel. Algorithms for detecting each of these six types of scene pixels are designed and used in the simulation. After the background model training phase, the background model can be updated when the new incoming scene frames appear. Thus, it can deal with a slow lighting change. In case of abrupt lighting change, either a system restarts with new background training or a larger intensity offset can be added to (or subtracted from) the background intensity range to allow a big shift in the pixel intensity value. In the computer experiments we use the ample variety of the data sets downloaded or prepared by our laboratory to test our new scene change detection method. We give the intermediate results as well as the final results. For performance evaluation we use the measure such as Precision, Recall, and F-Measure as performance indices. Overall speaking, our method performs well and can handle more problems such as the waking and sleeping persons than the existing methods. In additions, our method can be quickly adapted to the scene with fewer frames for model training.