Tracking of Human Activities

The problem of human body tracking

Difficulties:
1. Complexity and variability of the appearance of the human body
2. Nonlinearity and complexity of human motion
3. Lack of sufficient image cues about 3D body pose
4. Self-occlusion of limbs and torso

3D model of the body

We decompose the human body into truncated cones and ellipsoids. The body parts are organized as a tree with an ordered chain structure to provide the kinematic model of the limbs (Figure (b)). The cross-section of each cone is elliptical so that it can approximate torso and limb shapes more closely. These geometric solids are represented using quadratic equations, and correspondingly facilitate the computation of shape operators. The motions of the limbs are the rotations at the joints, and are represented using the relative rotation between local coordinate systems (Figure (c)).

**Figure:** Shape and kinematic model of a human body: The human body is decomposed into truncated cones and ellipsoids, and the joint motion is represented using rotation of the local coordinate system
$\begin{figure} \center{ \leavevmode \epsfxsize = 16.5cm \epsfbox{body.eps} } \end{figure}$

Measurement and state estimation

We use the same approach as in head tracking work. We first generate hypothesis about the size, pose, and color/intensity of each body parts and then project the corresponding 3D configuration to image plane to construct shape operators. The hypotheses are encoded into particles and they are propagated to the subsequent frames according to the filter responses, and the states are estimated. We use color/intensity filter (in addition to the shape filter for collecting edge responses) to make use of the color information and to alleviate the self-occlusion problem. The color of each body parts are estimated from the first one or two frames and a color/intensity filter is generated based on the depth order hypothesis about body parts.

Experiments on human walking

Experiments on synthetic data give a good tracking result as shown in Figure.

**Figure:** Synthetic walking sequence and tracked limb motion
$\begin{figure} \center{ \leavevmode \epsfxsize = 17.5cm \epsfbox{synthwalk.eps} } \end{figure}$

The tests on real video sequence are under way, and showing promising results.

**Figure:** Tracked body parts in treadmill walking sequence
$\begin{figure} \center{ \leavevmode \epsfxsize = 13.5cm \epsfbox{treadmill3.eps} } \end{figure}$

**Figure:** Tracked body parts in a walking sequence

About this document ...

Tracking of Human Activities

This document was generated using the LaTeX2HTML translator Version 2K.1beta (1.48)

The command line arguments were:
latex2html -html_version 3.2[math] -split 0 trac_body_html.tex

The translation was initiated by Hankyu Moon on 2001-02-20

Hankyu Moon 2001-02-20