3D Tracking for Gait Characterization and Recognition

Most current gait analysis algorithms rely on appearance-based methods that do not explicitly take into account the 3--D nature of the motion. In this work, we propose an approach that relies on robust 3--D tracking and has the potential to overcome the limitations of appearance-based approaches, such as their sensitivity to occlusions and changes in the direction of motion.

In previous work we have shown that using motion models based on Principal Component Analysis (PCA) lets us formulate the tracking problem as one of minimizing differentiable objective functions whose state variables are the PCA weights. Furthermore, the differential structure of these objective functions is rich enough to take advantage of standard deterministic optimization methods, whose computational requirements are much smaller than those of probabilistic ones and can nevertheless yield very good results even in difficult situations.

These methods can also be used to recognize people and characterize their motion. More specifically, we first used an optical motion catpure system and a treadmill to build a database of walking motions for a few subjects. We then captured both the motion of these subjects and of other people by running our PCA-based tracker on low-resolution stereo data. The resulting weights can then be used to recognize the people in the database and to characterize the motion of those who are not.

Figure 1: Clustering behavior of teh first PCA components

Because our tracking algorithm is robust to occlusions and insensitive to changes in direction of motion, our proposed approach has the potential to overcome some of the main limitations of current gait analysis methods. This is important if biometrics, defined as a measure taken from a living person and used as a method of identity verification or recognition, are to move out of the laboratory and into the real world.

1. Characterization

Only a few coefficients are used for tracking purposes since they encounter a big percentage of the database.

Figure 2: Percentage of the database that can be generated as a function of the number of eigenvectors.

Sequences captured using a low resolution and low framerate are used to show the performances of the system. The sequences have been tracked and the coefficients are recovered. This coefficients serves as motion characterizations since they encode the difference from the average motion.

Figure 3: Tracking results for subjects whose motion has been recorded in the database.

Moreover the fourth coefficient evolves almost monotonically with the speed, which will allow us to evaluate the speed of the walking subject. This can be recovered in a cualitative way, as shown in Figure ?? for the 3 tracked examples.

Figure 4: Fourth PCA coefficient as a function of the speed.

2. Recognition

The first coefficients tend to cluster, as shown in Figure1, since our approch handles the motion sequences as a whole, as opposed to a set of individual poses, thus making the clusters compact enough for direct classification. We use a simple k-means algorithm with the Mahalanobis distance as the similarity distance for recognition.

Figure 5: Recognition from stereo data.

The sequences from people whose motion has been recorded in the database fall into the cluster fromed by their recorded motions. Experimental results show that the first coefficients encodes general characteristics as weight, height, gender or age, and the following ones can be use to dishtinguish among people who share this characteristics.

PUBLICATIONS

R. Urtasun, P. Fua
3D Tracking for Gait Characterization and Recognition
In Proceeding of the 6th International Conference on Automatic Face and Gesture Recognition, Seoul, Korea, May 2004. IEEE Computer Society.