next up previous
Next: References

Gait Classification

Gait classification and gait analysis are applications of the analysis of human motion. Gait classification is the recognition of different kinds of gait, for example walking, running, or limping. Examples of these different kinds of gait can be seen in 1. The motivation of gait analysis is mainly seen in medical application. In gait labs, gait analysis is used for diagnosis of several diseases. This analysis is done by markers which are attached to the patient at the joints and limbs, in order to determine the position of the body parts. The fixing of the markers is very time consuming, but the method produces detailed information on the human movement. Other applications on gait analysis are found in rehabilitation and sport, or recognizing people by their gait for surveillance.

  
Figure 1: Different kinds of gait

One difference of gait classification in real world environment is that the clothing does not underlie any assumption. People wear shirts and trousers of any colour, some wear shoes, other walk barefoot.

First, the person has to be found in the image plane, which is done by background images or difference images, assuming a static camera. Then suitable features are extracted. The features used for gait classification has to fulfill several requirements. They have to describe the motion of the person, of course. This is either done by explicit modeling the body parts and analyzing the trajectories, or by extracting features directly from the images, the difference images or flow field.

Finding body parts

To extract exact trajectories of body parts, markers are attached to persons, or the images can be segmented manually. An example of a trajectory is shown in 2. From these trajectories, features have to be extracted. The most meaningful features describing the motion of body parts is the translation and the rotation angles, for example of the hip or the knee.

  
Figure 2: Tracked points at hip, knee and foot, trajectory of the foot (enlarged)

To track human body parts without using markers, an explicit model of the body is used. The detection of body parts from a single image without any knowledge about the posture is difficult, due to deformations, and occlusion. The legs and arms of someone walking are mostly occluded partially, sometimes even totally. The human body can be modeled by geometric figures, like cylinders, cones, ellipsoids, and others. There are several approaches, one [3] is shown in 3. The location of the body parts is done in a special range, starting with the most stable part. In most applications, this is the head. Using constraints for the other body parts, the localization of the other parts can be afterwards.

  
Figure 3: Example of a model representing the human body by cylinders

Extracting motion information from the flow field

Another possibility of extracting feature vectors for motion recognition is to use the optical flow field [2]. This field describes the motion in the image plane. It should not be too smooth, because different body parts move into different directions, as for example the arm and the trunk. The features extracted from the field should be person--independent, but distinguish between different kinds of gait. Suitable features extracted from the field are for example the mean or direction in several regions, or moments. In 4 the flow field of a walking person extracted by monotony operators [1] is shown. The feature vector consists of the mean of the flow in x-- and y--direction, and , and the acceleration . This method of feature extraction is less detailed than considering body parts, but it is less sensitive to errors by segmentation.

  
Figure 4: Features from the optical flow field

Classification

Having extracted the feature vectors, the classification itself is done. Learning algorithms are used for this item, as gait of different people vary a lot. It is possible to recognize someone by his gait, and even if the same person is not walking in exactly the same way every time. Classification can be done by dynamic programming, neural networks or by hidden Markov models (HMMs) [2], which is a common tool for speech recognition. It is also used in gesture recognition and in sign language. Gait is modeled as a sequence of states , each state is represented by an output probability which is the probability of observing the feature vector in state . The topology of the HMM should consider the fact that gait is a periodic action, one gait cycle contains a left and a right step. For each kind of gait one HMM with the initial state probability , the state transition matrix and the output probability matrix is trained, the classification is done by computing the HMM with the highest probability .





next up previous
Next: References



Bob Fisher
Tue Mar 30 16:01:12 BST 1999