中文摘要
視訊畫面之前景物體切割對於在視訊監控、辨識、檢索以及編碼等方面一直是重要的一環，針對移動鏡頭所進行之切割相較於固定鏡頭會有更廣泛的應用。由於視訊資料量龐大，在儲存及傳輸前皆需經過壓縮處理，直接利用壓縮資訊對移動鏡頭視訊內容進行移動物體切割便可省去對於原始影像之處理，為一適合即時應用之演算法。
近年視訊壓縮標準不斷演進，可用以進行切割的資訊也隨之改變。H.264標準中除了傳統的移動向量，還多出了七種模式的區塊分割可提供額外資訊。在我們所提出的移動鏡頭前景切割演算法中，首先利用運動向量以最小平方法求得背景相對於鏡頭之移動，再將與背景移動差異較大之區塊標記為前景區塊。在上述過程中我們利用區塊模式的不同，可以在逼近背景相對運動時取得較具參考價值之運動向量，並且在決定前景區塊時提升判斷之精確度，再加以我們設計之低通濾波器進行空間域與時間域之雜訊消除，最後以可調適性的門檻值來決定切割出的前景區塊。
利用移動向量能即時處理移動鏡頭之視訊內容，較許多空間域前景切割方法有更高之實用性。實驗結果顯示利用區塊模式資訊幫助前景分割，相較於只用移動向量可獲得較準確之結果，並且運算量低，可應用於即時系統中。
Foreground segmentation for video frames plays an important role in video
surveillance, pattern recognition, video indexing, and video coding. Due to the
large amount of video data, videos have to be compressed before storage and
transmission. Foreground segmentation based on the compressed information
prevents from processing the original frame, therefore, is an algorithm
suitable for real-time systems.
In recent years, video compression standards had been promoted rapidly. In
the H.264/AVC video coding standard, in addition to motion vectors, there are
also seven-mode block partitions which can provide extra information for
segmentation. In the proposed algorithm for moving object segmentation of video
acquired by moving cameras, we first approximate the relative global motion
model using motion vectors, then mark the blocks with motion vectors differed
from the global motion by a certain amount as foreground blocks. During the
procedure described above, according to the different block partition modes, we
can choose preferable MVs for estimating the global motion, moreover, improve
the accuracy of the judgments on blocks as being foreground or background.
Finally, we refine the results with spatial and temporal filters for removing
noises and segment foreground blocks by the proposed adaptive threshold.
With the use of motion vectors, we can process the video data for moving
cameras in the compressed domain. It makes this algorithm more practical in
applications rather than many object segmentation methods using spatial domain
information. It also reduces the computational costs and can be used for
real-time systems.