Skeleton Segmentation Overview


All silhouettes are assumed to be standing upright.

Points of interest (POIs) are defined as pixels or sets of pixels in a skeleton located at the endpoint of a segment. A POI may also be the intersection between two or more segments.

Purpose of POIs

Bone Properties

Problems

Method Overview

Alarms

Results


Purpose of tracking POIs and segments

Identify body parts and arm activity
Improve skeletons
Improve "blobs"


Bone Properties

Torso

Head

Arms

Legs

Some common configurations:


A

B

C

  1. head, endpoint & intersection at top, near-vertical; arm, endpoint & intersection with torso, non-vertical; torso, 2 intersections, long, high slope; legs, intersection & endpoint, high slope
  2. torso and vestigial legs only; torso includes endpoint
  3. torso only (head can be created as part of segment)

Problems

Loops: 2 segments with the same endpoints--2 intersections


Bone Spurs: short or illogically located bone


Method Overview

First determine the POIs and segments of the skeleton.

Find the head based on the silhouette.

  • Highest unchecked pixel
  • Connected components downward
  • Centermost above threshhold is head
  • Dimensions may offer shoulder guess

Find the end of the head using three different calculations:

Width-Based
Gradient-Based
Circle-Fit
Sample Image and Video:

Head-based identification.
  • Head: reliable starting point (relatively stable, easily found)
  • Torso: top attached to head
  • Legs: tops attached to torso
  • Arms: highly mobile, therefore highly variable
Arms must be addressed further.
  • Two levels of certainty to keep arms during appearance and rotation.

  • Broken arms adjusted to simplify skeleton and improve accuracy.

Also, the height:width ratio is calculated, and if it is
greater than 3.9, legs are possible.
greater than 3.0, legs may be possible, based on history.
less than 3.0, legs are not possible unless history strongly contradicts.
This helps avoid marking arms in a partial silhouette as legs and helps narrow the search for arms.



Alarms

Arm positions are analyzed based on slope. The first and lowest alarm is the horizontal arm position. This is the lowest level because when an arm first becomes distinguishable from the torso, it almost always falls into this category. It is identified by an orange bar across the top (for left arm) or bottom (for right arm) of the image.

The second alarm level is the vertical arm position. An arm will move from the first alarm level to no alarm before moving into this position, although this is not yet checked by the program. A red bar down the left or right side of the image indicates a raised arm.

Globally, the highest alarm is reached when there are multiple silhouettes and both level and raised arm alarms.



Results

August 2002

Sequence # Frames Successful ID
walaim1 68 94.12%
walaim2 115 93.04%
walup1 104 87.5%
walup2 97 93.81%
mikeaim 512 92.77%
mikemulti 548 91.24%
TOTAL 1444 91.96%

These sequences are of a single person fully visible outdoors. The "wal" sequences were of a quick mover late on an overcast afternoon in front of a wall and moving branches. The "mike" sequences were of a slower mover who rotated after walking to the center of the picture; the setting was a bright, early afternoon on grass in front of a pale wall.

Sample Videos


| Project main page


jdever@cs.ucf.edu

Last updated: 23 NOV 2005
Created: 9 OCT 2001