Human posture estimation method from a single image using genetic algorithm and fuzzy inference

In this paper, we propose a method for estimating the observation direction of a person by applying fuzzy inference to the features obtained when a genetic algorithm is used to recognizee human posture from a natural image. Posture recognition uses the constraints obtained by representing the human body as parts expressed as connected structures to reduce the search space of the genetic algorithm. First, the head, which is a relatively easy part of the body to detect, is detected by pattern matching. Then, based on the position of the head, the body, arms, and legs are detected successively. A complex matching model becomes unnecessary because each part is approximated by a combination of straight lines and ellipses. The generated image of each part constructed from straight lines and ellipses is compared to the input image and optimized to obtain some level of overlap by the genetic algorithm. The obtained features are input to a fuzzy inference neural network, and the observation direction of the person can be estimated. The fuzzy rules can be automatically extracted by using Kohonen's self-organizing algorithm and the minimum least squares algorithm. The method of this paper applies to many people, and its effectiveness is verified.