Event Understanding of Human-Object Interaction: Object Movement Detection via Stable Changes - Intelligent Video Event Analysis and Understanding - page 204

Input Image Partial Frame Subtraction

Foreground Region

Compare

(b)

(a)

Fig. 5 Partial frame subtraction. (a) An overview of partial frame subtraction. (b) Motion
detection result via partial frame subtraction. The left images are the input images, the center
image is the partial frame subtraction result, and the right images represent the foreground
regions. In the center images of (b), rectangles are overlaid on regions where the method de-
tects motion: the outer rectangles represent that the method detects the foreground's region is
moving via the partial frame subtraction, the inner rectangles are the partial regions which the
method detects via the proposed method. When you focus on the human region, the contours
of the regions are almost the same, but the method detects movement of human arm by the
partial frame subtraction.

2. Update-prohibited region (object state)
If the stability value of the region in the transition region is over some threshold
S th ( S S th ), it transfers to the update-prohibited region. This attribute repre-

sents the region is non-object because the region was moving for a long time. To
avoid classifying long-term change caused by non-objects as placed objects, the
update-prohibited regions transfer only to the update-prohibited regions.
3. Static region (non-object state)
If the stability value of the region with the transition region is S

( t ) >

S th ,it
transfers to the static region. In the background update stage, static regions are
deleted from the foreground region list and are inserted to the layered back-
ground.

( t ) < −

In our implementation, we set the threshold parameter S th = 20.

4.2 Stable-Change Classification

After the method extracts static regions, the method rejects the static regions with-
out object movements (e.g. shadows, small object shift), then the method classify
the object regions into object placement and object removal. To reject the stable
changes without object movements, we set some threshold parameters - Bhattachar-
rya distance between HSV histogram (we employed Perez's HSV histogram [18] in
our method) of the input image and the background image in the extracted regions