Event Understanding of Human-Object Interaction: Object Movement Detection via Stable Changes - Intelligent Video Event Analysis and Understanding - page 209

proposed method works correctly. At the same time, the size of the remote con-
trol is small (only 12 ), the system successfully detects object movements.

Fig. 10 shows the detection result of the sequence when a person places a box on the
table, sits down on the chair and then removes the box. In this case, even the person
is sitting down and his body is occluded by a chair, the method does not detect the
person as an object but detects object removal. So, the motion-based classification
method of object and non-object works robustly.
The average process time of the proposed method was 130[ms/frame], so the
proposed method works in sufficient frame-rates.

× 5 [ pixel ]

5.2 Limitations

The proposed method has some limitations. First, the background subtraction
method can't handle strong illumination changes (e.g. switching off the lights or
opening a curtain). Second, the proposed method can't handle movement of furni-
ture (e.g. opening and closing of a door, large shift of chair) because movement of
furniture is neither of “object placement” and “object removal”, that the proposed
method can handle. Third, the proposed method hardly detect object movements
when the color of the object is very similar to near object's (e.g. when a white tele-
phone is placed near white curtains). In this case, robust object movement detection
is difficult by using images captured only in a single viewpoint. So, to detect objects
more robustly, the method needs to integrate multiple viewpoint information.

6Con lu ion

This chapter proposed an object movement detection method in a household envi-
ronment via the stable changes of images. To detect object placements and object
removals robustly, the method employs the layered background model and the edge
subtraction based classification method. In addition, to classify objects and non-
objects robustly though the changed regions are occluded, the method uses motion
history of the regions.
Our experiment shows the method detects objects robustly even though the object
size on the image is small, and the method distinguishes between object placement
and removal appropriately. The method runs in sufficient frame rates, so the pro-
posed method works well.
Future tasks are handling large object shift such as movement of furniture, inte-
grating multiple viewpoints to detect objects more robustly.

References

1. Gupta, A., Kembhavi, A., Davis, L.: Observing human-object interactions: using spatial
and functional compatibility for recognition. IEEE Transactions on Pattern Analysis and
Machine Intelligence 31(10), 1775-1789 (2009)
2. Navneet, D., Triggs, B.: Histograms of oriented gradients for human detection. In: CVPR
(2005)