Gait Recognition using Dynamic Afﬁne Invariants
We present a method for recognizing classes of human gaits from video sequences. We propose a novel image based representation for human gaits. At any instance of time a gait is represented by a vector of afﬁne invariant moments. The invariants are computed on the binary silhouettes corresponding to the moving body. We represent the time trajectories of the afﬁne moment invariant vector as the output of a linear dynamical system driven by white noise. The problem of gait classiﬁcation is then reduced to formulating distances and performing recognition in the space of linear dynamical systems. Results demonstrating the discriminate power of the proposed methods are discussed at the end. We live in a dynamic world, constantly analyzing and parsing time varying streams of sensory information. Almost all biological creatures equipped with the sense of vision use dynamic cues to analyze their surrounding for critical survival decisions. Clearly there is an abundance of information embedded in the dynamics of visual signals In this . work we focus on extracting and exploiting the temporal structure in video sequences for the purpose of recognizing human gaits.
Observing a person walking from a distance, we can often tell whether the subject is a human, identify their gender, or make predictions about individual traits like age or physical health. We postulate that such information is encoded not necessarily in the static appearance, but mostly in the dynamics of the moving body. In Johansson’s experiments one cannot tell much from a single frame, however when the sequence is animated suddenly the scene is easily parsed. Johansson’s experiments show that even in the lack of all comprehensible static content, the dynamics of a few moving dots can contain sufﬁcient information to 1Naturally there is also a great deal of information in the photometry and geometry of the scene that can be conveyed in a single static frame. However, in this study we concentrate on the scene dynamics. correctly decipher the underlying physical phenomenon. In this paper we address the problem of recognizing a person walking from one jumping, runnning, hopping or dancing, independent on the person and his pose. The problem of image-based human motion analysis and recognition has been receiving considerable attention in the literature. Most of the proposed approaches involve tracking the pose of the human body, represented either as kinematic chain of body parts , or as spatial arrangement of blobs or point features . Statistical models, such as standard and parametric Hidden Markov Models are then ﬁtted to the tracking data and likelihood tests are used for recognition. In mixedstate statistical models for the representation of motion have been proposed, and in particle ﬁlters have been applied in this framework for estimation and recognition. In linear gaussian models have been used, and recognition is performed by deﬁning a metric on the space of models