Tracking capture in video motion.

Baritz, Mihaela ; Cristea, Luciana

1. INTRODUCTION

Motion capture of human body can be an effective method of creating realistic human motion for animation or for different technical or medical studies. Unfortunately, the quality demands for animation place challenging demands on a capture and visualization system. The capture solutions that need these demands have required specialized hardware that is invasive and expensive solutions. For that computer vision could make animation data much easier that another method and in plus it can obtain information's to use them in different fields. In this paper it is presented a methodology for analyzing the 2D motion given image observations; use this as a tool for understanding the problem and for survey the evolution of a human body in different moments of walking process.

Synthetic experiments confirm that these situations would arise in practice. The experiments show how even simple visual tracking information can be used to understand the sort of gait for different people or to create a standard human motion but even with video tracking. (Bregler C., Malik, J. 2002)

Motion capture is an integrate method for creating the movement for computer animation or for different technical or medical studies, but motion capture also has its share of weaknesses.

The development of new and improved methods of editing and processing motion capture data has made great strides in making motion capture a more viable tool for animation production or for understanding biological process, technical and medical structures.

Another problem of motion capture has been the practical challenges of acquiring data. While research has made progress on using the data, capture techniques have evolved slowly, having problems with inertial time of acquisition and visualization. In special tracking technologies, based either on mechanical or magnetic sensors, or specially designed cameras viewing light markers, are required to create the observations that are processed into motion data. While these systems have improved in their dimensions, reliability, precision, and range, they are still generally expensive and some times hard to be used. These activities of motion capture must be performed by dedicated places providing specific environments for the recordings. Any "performer" could be captured in any setting that is desired. Using standard, high speed or thermal video cameras and special conditions for recording is an integrated structure as it could meet these goals.

The use of a single camera is a particularly situation only for static position and small or medium movements.

[FIGURE 1 OMITTED]

It offers the lowest cost, simplified setup, the potential use of natural or artificial light sources and a space without any modifications. The creation of motion capture data from a single video stream seems like a plausible idea but only for small and medium movements (short distance).

2. THE MOTION CAPTURE PROBLEM

The goal of motion capture is to record the movement of a "performer" (typically, but not always, human) in a compact, usable, repetitive manner and to be possible to rebuild these movements.

For this paper, we are concerned with the gross motion of the body, on the short distance (3 m), because the specific capture of facial or hands motion poses a different set of problems like skin color, 3D images and vibration stability. In computer graphics/computer vision studies usually the human body is divided into a small number of rigid segments that rotate relative to one another. This approximation is simply because the human knees, elbows and ankles do not have a single pivot point. The true motions of more complex joints are sometimes examples of kinematics approximations. The motion capture problem we consider, therefore, must have the following form: given a single stream of video observations of a people with normal/disabilities gait, analyze the tracking of each point, and compute a 3D skeletal representation of the motion using sufficient quality to be useful for animation. But for medical studies of the movement of the human body, like in Parkinson disease, are sufficient to use the rigid form of the human body parts. (Gleicher M., Ferrier, N 2003)

3. EXPERIMENTAL SETUP

Virtual humans are articulated figures modeled with multiple layers: a virtual skin is usually attached to an underlying skeleton, which animates the whole body. The skeleton is a hierarchically organized set of joints, and this set depends on the animation requirements and the fields of applications. To create morphologically correct skeletons it will be better but this can turn out to be quite costly. Real humans have so many degrees of freedom that virtual models frequently are necessary to omit some of them or to minimize the importance into movement actions.

[FIGURE 2 OMITTED]

The capture problem is sometimes difficult: the articulated model does not accurately reflect the real people, articulations lead to self-occlusions, even the articulated models contains many degrees of freedom, the skeleton is internal and therefore cannot be observed directly, also the clothes are an important part into the observations and recordings the movement's images. Our information sources are 2D-bidimensional and some occlusion is possible in time of recordings (like hands passing in front of hip joints). In addition, the optical medium for recordings provides a finite resolution (spatially and temporally), and the parameters of real cameras are difficult to obtain precisely (light, vibration, humidity). Having these limitations, it is not a problem that the practical approach to motion capture for moving tracking of the human body involves can be used in this method without these limitations. For example, if we observe a point on the people in an image created by a camera, we cannot determine the position of the point; only constrain its location to lie along a ray. For that, we assume an idealized pinhole camera model such that the ray is defined by the camera's focal point and the point on the image plane. In practice, a video camera has a finite resolution so observations are only localized to a region of the image plane and the space of human body movement must be in this region. (Monzani, J.S., 2002)

Additional information is needed to determine the position of a point in space, like axis system and calibration. A variety of sources can be utilized in various computer vision techniques, and a few can be applied to motion reconstruction. Most methods assume strong models to place further restrictions on possible poses. There are a wide range of visual tracking techniques in the practice ranging from edge feature based to region based tracking, and brute-force search methods to differential approaches. Edge feature based tracking techniques usually require clean and good data with high contrast object boundaries. But on human bodies such features are very difficult to obtain; clothes have many folds, environmental light is sometimes inconstant and also the images background has different colors in different directions.

Also, if the left and right leg has the same color and they overlap, they are separated only by low contrast boundaries.

Region based techniques can track objects with arbitrary texture and attempt to match areas between consecutive frames.

[FIGURE 3 OMITTED]

[FIGURE 4 OMITTED]

This requires the estimation of 6 free parameters that describe this deformation (x/y translation, x/y scaling, rotation, and shear).

4. RESULTS AND CONCLUSIONS

In our researches we used two video cameras, one of them withhigh speed and other with normally speed, to record the human body movements between initial point and end point situated at 3 meters distance. Initial the people wearing black cloth walks with a normal speed (also measuring the ground forces develop in gait process by a force plate) on this distance and we're recording the image of all points attached on the legs joints. Using these trajectories of the leg joints we can transfer to a human virtual model to analyze the movements, the joints positions in different movement moments and also to establish a data base for each people. (www.adobe.com)

Also, these recordings are possible to be used to simulate a movement and to establish the limits or to calculate the forces develop in muscles or joints during gait process.

[FIGURE 5 OMITTED]

5. ACKNOWLEDGMENT

Researches are part of Grant A1088 with CNCSIS Romania and we've developed the investigations with apparatus from Research Platform "SAVAT", University Transylvania Brasov.

6. REFERENCES

Bregler C., Malik, J. 2002, Video Motion Capture, UCB/CSD 97-973;

Gleicher M., Ferrier, N 2003, Evaluating Video-Based Motion Capture;

Monzani, J.S., 2002, An architecture for behavioral animation of virtual humans Ecole Politehnique Federale de Laussanne, Suisse;

www.adobe.com, accessed 2008-06-15.