Towards Effective Human Motion Descriptors
Jakub Valcik
Supervisor: Pavel Zezula
Consultant: Jan Sedmidubsky
NOV 18, 2013

Outline
¨Motion Capture Data
¨Motivation
¨Evaluation Methodology
¨Results
¨Summary
2
cartwheel.gif

Motion Capture Data
3
¨Digitalized human motion
¤joint coordinates,
¤ euler angles
Movement
MoCap Data
Features
¨Features
¨Derived descriptors

>

Optical Systems
4
¤Triangulations of the 3D position from image sensors data
¤Multiple high-speed video cameras 2~48 (even 300)
¤Markers or surface features
¤Passive markers
nRetro-reflective material reflects light generated near the camera lens
n+Wireless -Marker swapping
¤Active markers
nLED emitting own light
nMarker identified by modulation of amplitude, pulse width, time window
¤
¨

Optical Systems, cont.
5
¨Passive imperceptible markers
¤Up side down approach
¤Photosensitive markers
nDepth Map
¨Markerless
¤Analysis of video
nIdentify human forms and brake down into constituent parts for tracking
¤Stanford, UMD, MIT, MPI
¤MS Kinect, Asus Xtion, PrimeSense Carmine, Organic Motion, Xsens

Non-optical systems
6
¨Inertial systems
¤Miniature inertial sensors
¤Wireless comunication
¤Position error accumulates over time
¤Wii controller
¨Mechanical systems
¤Exo-skeleton system tracks angles directly
¨Magnetic systems
¤Relative intensity of the voltage or current of coils

Applications
7
¨Health care – success of rehabilitative treatments
¤Range of joint angle rotation
¨Sports – performance aspect comparison
¤Variability of same motion – pole-jump, figure skating
¨Security – person identification, event detection
¤Gait recognition, stealing, fighting
¤Home for the elderly
¨Computer animation – realistic motion synthesis
¤Motion retrieval

Premise
8
¨Ultimate descriptor solving problems of all applications does not exist

Purpose
9
¨Application => Purpose
¨Action oriented
¤What?
¤Action, style of action, event detection
¤Logically similar movements
¨Subject oriented
¤Who?
¤Subject recognition, age, gender, drunkenness, pregnancy, skeletal disease

\\afrodita\home\xvalcik\_profile\Desktop\key-pose-joint-angle-extraction.png
Similarity Model
10
¨Pose – skeleton configuration in one frame
¨Pose features – extracted from one pose
¤Distances between joints/planes, joint angles, velocities, accelerations, powers, torques,
directions
¤Optional quantization
nrelational features, fuzzy features
¨
¨Distance between:
¤Poses – LP, Hamming, Mahalonobis
¤Sequences – DTW, UTW, Uniform scaling

Evaluation Methodology
11
¨Choose optimal descriptor for given purpose
¨
¨
Similarity Models
Joint Angles
L1
UTW
Muller Features
Hamming distance
DTW
Torso Declination
L1
UTW
Purpose
winner.jpg (400×300)
How to interpret purpose?
Test Dataset
How to chose optimal model

Ground Truth as a Purpose
12
C:\Users\Kuba\Documents\Cropper Captures\jump.gif C:\Users\Kuba\Documents\Cropper Captures\bend.gif
C:\Users\Kuba\Documents\Cropper Captures\throw.gif C:\Users\Kuba\Documents\Cropper
Captures\hopOneLeg.gif

Measures
13
¨Retrieval Oriented
¤Mean Average Precision (MAP)
¤Mean Reciprocal Rank (MRR)
¤Discounted Cumulative Gain (DCG)
¤K-Nearest Neighbors
¨Space Oriented
¤Dunn index
¤Davis-Bouldin index
¤Distance Distribution
¨Time consumption

Results
14
¨Dataset : HDM, CMU

Summary
15
¨Purpose oriented descriptor evaluation
¨Purpose represented as a ground truth
¨Provided measures
¨
¨Future work
¤Distance distribution
¤Statistical testing
¨

Towards effective human motion descriptors
16
¨Q & A
¨
¨Thank you for your attention