An Efficient Method to Detect People Carrying Object.
We have designed a silhouette-based method to determine whether or not
a person is carrying an object. It combines two basic observations:
Human body shape is symmetric, and people exhibit periodic motion while
they are moving. During tracking, the people's periodic motion of people
is estimated, and the regions on the silhouette which violate the symmetry-constraints
are determined. Those results are combined to determine if a person
is carrying and object, and to segment the object from the silhouette.
In cooperation with Hydra during tracking, we construct an appearance model
for each
carried object, so that when people exchange objects, we
can detect "who" carries "which" object via an analysis
of the segmentation.
Publications :
-
An Efficient Method to Detect People Who are Carrying
Objects,
I. Haritaoglu, D. Harwood, L. Davis,
(In Preparation) Will be Seventh Interactional Conference on Computer
Vision, 1999
Silhouette Analysis:
-
Extraction of silhouette edges
-
Detection of possible body part location head on silhouette boundary
edges
-
Corner points
-
Convex hull vertices(Primary/secondary)
-
Vertical and Horizontal projection histogram analysis of binary foreground
region.
Symmetry Analysis:
-
Human body image is symmetric while standing (walking
and running)
-
balance on feet, around a symmetry axis
-
Body is segmented according to symmetry
-
Regions which obey the symmetry constrains (blue)
-
Regions which violate the symmetry constrains (orange)
-
Legs, arms may violate the symmetry periodically
-
big objects carried by human violates symmetry continuously
Periodic Motion :
-
Periodic Motion is determined using shape changes over time
-
X and Y Projection Histograms of binary silhouette
-
Similarity plots for X and Y projections is calculated using last N silhouettes.
(N=50)
-
Row base auto-correlation applied to similarity plots for periodicity
calculation
Symmetry + Periodicity:
-
Periodic motion analysis is applied to Non-Symmetric regions.
-
Regions which has not periodic motion are classified as objects
carried by person (red regions)
-
Textural templates of objects are generated and updated as long as they
are detected
Results:
Movies:
