Stereo Vision Object Recognition Through Training

The objective of this computer vision project is to recognize objects in a robot's environment, using stereo vision, computer vision, digital signal processing, and a hueristic algorithm with weighted characteristics through training a model.Many researchers have tried to solve the problem of computer vision object recognition by complex methods. Many of these methods cannot be used for real-time object recognition for a robot because the object representation is too complex and the calculations take too long. In this thesis I present a simple method for object recognition. My work is based on various other methods that use region growing to separate the object from the background, and I represent the object as a feature vector of global invariant shape characteristics, such as spatial moments of inertia. However, unlike previous methods that use binarization, I retain the grayscale value of the object before region growing and use this information to limit the search when finding the object. The method trains an object by repeatedly gathering shape information of the object in different positions and different distances. This, along with the grayscale value of the trained object is then used to find the object with the closest numerical match. Stereo vision is used to find the world coordinates of an object so that a robot may move to the object, but shape information for left and right camera views are not combined because the problem of calculating volume from just two images is not possible. Since the search is reduced to objects that are of certain grayscale values, and the representation of the object is just a feature vector, the computations is very fast. Therefore, this method is suitable for real-time recognition of simple objects for the use by a robot.All computer vision source code is included (written in C )