Scene interpretation and object recognition for mobile robots equipped with cameras

Sammanfattning: A method consisting of separate steps for object detection and object recognition has been developed for use with digital camera images. The detection step uses double-opponent maps and the Hough transform to find likely object candidates without having to search the entire image. The recognition step is built around a modular neural network (MNN) trained using backpropagation into which colour histograms and eigenimage projection coefficients are fed as main cues along with simple region information. The detection approach was found to be 45 times faster than an equivalent brute force approach and the resulting algorithm is found to be fast enough in order to allow use in real-time robot landmarknavigation.Substantial differences in recognition ability between different types of objects were found when applied on images from a typical office environment. The performanceof the algorithm indicates that visual object recognition can be achieved in real-time with only off-the-shelf equipment.