MODEL BASED POSE ESTIMATION USING IMAGE PROCESSING

Abstract

MODEL BASED POSE ESTIMATION USING IMAGE PROCESSING By ABDUL WAHID Pose estimation is considered an important component in many pattern recognition and computer vision systems. One of the important applications of pose estimation is in model based object recognition. In pose estimation, the problem is to determine the orientation and position of an object which would result in the projection of a given set of three dimensional points into a given set of image. The thesis presents general method for pose estimation by fitting model of object with arbitrary curved surfaces on the image of the object. A model of an object of known dimensions is created. Simple models are generated in MATLAB and complex models in Pro-e. Computer vision and image processing techniques are applied to compare the 2D projection of the model with the image taken of that object. Cost function, which is the mean square of difference between the images taken from the camera and the 2D projection of the model, is calculated. Various search techniques like stochastic gradient and genetic algorithms are used to find minima of cost function. The required point will give the desired pose of the object. When the cost function is smooth the stochastic gradient method is the ideal one but in our case the cost function has too many local minima thus the genetic algorithm was used iv to get acceptably good solution. Considerable attention has been given to issue of robustness and efficiency and the technique should serve as a practical basis for model fitting in most applications of model based vision.