ON THE MOVE - utwente.nl€¦ · Body pose and movement tells us about: Action / Intention Affect /...
Transcript of ON THE MOVE - utwente.nl€¦ · Body pose and movement tells us about: Action / Intention Affect /...
ON THE MOVE AUTOMATICALLY MEASURING BODY MOVEMENT
RONALD POPPE
Human behavior can be:
Verbal
Nonverbal
Nonverbal behavior consists of:
Para-verbal (acoustics)
Facial expressions
Gaze
Gestures
Body movement
HUMAN BEHAVIOR
This talk is about:
Body pose
Body movement
And how to measure these automatically
Any questions: just ask straight away!
HUMAN BEHAVIOR
Theory:
Body movement
A short history of measuring body movement
Body movement representation
Practice:
Sensors and devices
Processing and analysis
The future
OUTLINE
THEORY: BODY MOVEMENT
BODY MOVEMENT: QUESTION
1 2 3
4
Q: who said
“No, it wasn’t
me honey”
Body pose and movement tells us about:
Action / Intention
Affect / Mental state
Attitude
Body movement can be conscious or unconscious (automatic)
Individual: frustration, aggression
Interaction: dominance, turn-taking and interruptions
BODY MOVEMENT
We can measure body motion to:
Answer hypotheses about nonverbal behavior itself
Answer hypotheses about phenomena with nonverbal correlates
Make (computational) models of human bodily behavior
Control interfaces and games
Top-down (hypothesis-driven) vs. bottom-up (data-driven)
BODY MOVEMENT
Measurement of body motion:
Qualitative (coding)
Quantitative (motion capture)
Qualitative is typically performed manually:
Have observers count/code specific behaviors
Analyze frequency of behaviors over time
BODY MOVEMENT
Coding of nonverbal behaviors:
Interpretation can be given
Irrelevant movements can be ignored
Coding scheme should be known beforehand (requires analysis)
Time-consuming (often multiple coders)
Subjectivity of the coders
Depends on viewpoint/quality of cameras
Qualitative, not quantitative (differences in speed, direction and
form should be additionally coded)
BODY MOVEMENT
Automatic measurement of nonverbal behaviors:
Quick (with minimal effort)
Objective
Quantitative
No interpretation of the behavior
Irrelevant behavior affects the analysis
BODY MOVEMENT
BODY MOVEMENT: HISTORY
Aristotle (330BC), Da Vinci (1500), Borelli (1680)
BODY MOVEMENT: HISTORY
Étienne-Jules Marey (1878) and Eadweard Muybridge (1880)
BODY MOVEMENT: HISTORY
Chronophotographic gun
BODY MOVEMENT: HISTORY
BODY MOVEMENT: HISTORY
BODY MOVEMENT: HISTORY
BODY MOVEMENT: HISTORY
Gunnar Johansson (1973)
BODY MOVEMENT: HISTORY
BODY MOVEMENT: REPRESENTATION
BODY MOVEMENT: REPRESENTATION
The body can be represented as segments and joints
Segments are body parts, and have a length
Joints connect segments
Movement takes place at the joints
Humans have 206 joints
Most of them in the back, hands and feet
BODY MOVEMENT: REPRESENTATION
All joints together form a kinematic tree
Two joints at either end of a segment are connected
The root joint is at the top of the tree
End-effectors do not have only
one attached segment
All connections from root to end-
effort are called a kinematic chain
Joints higher in the tree affect those
below in the chain
BODY MOVEMENT: REPRESENTATION
Joints can move in different directions
Each direction is called a degree of
freedom (DOF)
Joints can have up to 3 DOF
How many DOF in the elbow?
How many in each shoulder?
And in the knee?
BODY MOVEMENT: REPRESENTATION
Joints rotate around axes
For each axis, there is a feasible
range of motion
This range is bounded by rotation
constraints
The DOF, with constraints,
determine the space of movement
BODY MOVEMENT: REPRESENTATION
The order of applying the
rotations is important!
We can specify a body pose
by the rotations of all joints
BODY MOVEMENT: REPRESENTATION
Often, we are interested in the location of the end-effectors (hands,
feet, head)
The same location of a hand can have different sets of joint rotations
Comparing positions based on joint rotations is difficult!
Why not use joint locations directly?
BODY MOVEMENT: REPRESENTATION
BODY MOVEMENT: REPRESENTATION
Joint angles
• Easy to analyze single joint
• Difficult to interpret
• Difficult to calculate
distance between joints
Joint positions
• Difficult to analyze single
joint
• Easy to interpret
• Easy to calculate distance
between joints
Segment
lengths
Joint locations are points in 3D
space
Each point can be written as:
(x, y, z) with distances along axes
from the origin
The origin is the point (0, 0, 0)
Axes need to be defined
E.g. origin in center of a room, axis
in direction of walls
BODY MOVEMENT: REPRESENTATION
A body pose can be
described by the joint
locations of all joints
Joint positions are highly
correlated
Moving in a room causes all
joints to change
BODY MOVEMENT: REPRESENTATION
(2.0, 1.8, 3.9)
(1.7, 0.1, 4.3)
(2.8, 1.8, 0.8)
(2.5, 0.1, 1.2)
Global joint locations:
Locations depend on position in the room
Locations are relative to the origin in the room
Why not use the position in space + the local body pose?
Local joint locations:
Locations relative to “something” in the body local origin
BODY MOVEMENT: REPRESENTATION
Local origin is root joint
Often, the root is the pelvis
Location of a joint = location of pelvis +
relative location of joint
Local location = global location – origin
BODY MOVEMENT: REPRESENTATION
Joint locations no longer depend on position in space
We call this position normalization
We can now quickly compare poses and see if they are the same!
But what happens if you turn around?
BODY MOVEMENT: REPRESENTATION
The pose is exactly the same but…
Rotation changes global and local joint locations
Can we do normalization similar to what we did for position?
BODY MOVEMENT: REPRESENTATION
(0.2, 0.7, 0.0)
(0.0, 0.7, 0.2)
0.2
0.2
We can describe the joint locations relative to the
facing direction of the person
Facing direction is determined by the hips
We align the joints to one of the axes
So the facing direction always is the same after
rotation
We call this orientation normalization
BODY MOVEMENT: REPRESENTATION
First, we determine how much a person is rotated
We calculate the angle between the facing direction and the axis
Then we rotate all joints with this angle around the root joint
BODY MOVEMENT: REPRESENTATION
Recap:
We can represent body poses as 3D joint locations
We can normalize joint locations for global position
We can normalize joint locations for global orientation
We can now calculate differences between poses!
Needed for the quantitative analysis of body pose and motion
BODY MOVEMENT: REPRESENTATION
Poses are aligned on the root joint
Pose difference is the summed distance between all pairs of joints
BODY MOVEMENT: REPRESENTATION
What about calculation of motion?
Motion corresponds to body poses over time
So we can calculate the difference between subsequent frames!
By averaging distances over time, we can calculate average motion
For the whole body or per joint/limb
BODY MOVEMENT: REPRESENTATION
PRACTICE
Theory:
Body movement
A short history of measuring body movement
Body movement representation
Practice:
Sensors and devices
Processing and analysis
The future
OUTLINE
BODY MOTION: SENSORS AND DEVICES
Used extensively in movie and games
BODY MOTION: SENSORS AND DEVICES
With special devices:
Mechanical: measuring angles and distances directly
Marker-based: visible markers on the body
Inertial: magnetic sensors on the body
Without special devices:
Vision-based with depth camera
BODY MOTION: SENSORS AND DEVICES
Mechanical: Animazoo Gypsy 5
Suit with sensors that measure
angle and extension
Direct measurement
Multiple actors
Difficult to set up
Limited number of joints
Limits the freedom of movement
BODY MOTION: SENSORS AND DEVICES
Marker-based: Vicon MX
Retro-reflective markers attached to
a suit or straps
Passive or active
Multiple actors
Freedom in marker configuration
BODY MOTION: SENSORS AND DEVICES
Large recording environement required (many cameras)
Invisible markers due to occlusion
Markers might be swapped
Markers might fall off
BODY MOTION: SENSORS AND DEVICES
Inertial: Xsens MVN
Sensors with gyroscopes in straps
Large performance space
Outdoor recording possible
Multiple actors
Risk of drift due to metal in environment
BODY MOTION: SENSORS AND DEVICES
Depth camera: Microsoft Kinect,
Asus Xtion, PrimeSense
No sensors on the body
Multiple actors
Occlusion is a problem
Accuracy is relatively low
Difficulties with direct sunlight
Limited in performance space
BODY MOTION: SENSORS AND DEVICES
BODY MOTION: PROCESSING AND ANALYSIS
After recording, we have a lot of data
Now what?
BODY MOTION: PROCESSING AND ANALYSIS
The cook book:
1. Transform all data to column format
Each DOF is a column
Each time frame is a row
2. Select start and end frames, select corresponding rows
3. If there is noise, apply a filter
4. Normalize for position (global to local)
5. Normalize for orientation
6. Calculate dependent variable
BODY MOTION: PROCESSING AND ANALYSIS
Filtering of data
Remove outliers
Smoothen tracking inaccuracies
Median filter with small window length often does the trick
BODY MOTION: PROCESSING AND ANALYSIS
Dependent variables:
Amount of movement
Average distance to reference pose
For the whole body or for individual limbs, e.g.:
Total movement in the left leg
Difference in pose between left and right arm
BODY MOTION: PROCESSING AND ANALYSIS
Distance between two subjects
(Average) distance over time
More complex dependent variables
BODY MOTION: PROCESSING AND ANALYSIS
BODY MOTION: THE FUTURE
Bottom-up research:
Collect data
Mine patterns
Typical in information retrieval and pattern recognition
Why not in psychology?
BODY MOTION: THE FUTURE
More detailed/accurate body motion measurements
Facial expressions
BODY MOTION: THE FUTURE
In “the wild”
Deal with “non-cooperative” people
More natural behavior (also in actual environment)
Longer-term behavior
BODY MOTION: THE FUTURE
Recorded footage
Learning behavior from freely available recordings
Automatic re-analysis of material
Quantitative instead of qualitative
BODY MOTION: THE FUTURE
Experiments with virtual characters and robots
Control over behavior (modalities)
Analyze variability of behavior between subjects
BODY MOTION: THE FUTURE
Klette & Tee, “Understanding Human Motion: A Historic Review”,
Computational Imaging and Vision: Human Motion, 2008
Moeslund, Hilton, Krüger & Sigal (Eds.) “Visual Analysis of Humans -
Looking at People”, Springer 2011
Poppe, “Vision-based human motion analysis: An overview, Computer
Vision and Image Understanding”, 2007
Poppe, “A survey on vision-based human action recognition”, Image
and Vision Computing, 2010
Poppe, Van der Zee, Heylen & Taylor, “AMAB: Automated
Measurement and Analysis of Body Motion”, Behavior Research
Methods, forthcoming
BODY MOTION: SOME LITERATURE
Thanks!
Contributors: Dirk Heylen, Paul J. Taylor, Sophie van der Zee
Otherwise, send me an email: [email protected]
QUESTIONS!?