Abstract:A detection algorithm was developed to detect objects with large variances in color, texture, and shape. The algorithm does not need any manual selection or segmentation to automatically extract edge fragments from training samples to form a dictionary of templates. A modified statistical region merge segmentation algorithm and the Adaboost learning strategy are used to dramatically reduce the effect of noise caused by the automatic extraction of the edge fragments. The matching cost used in each weak classifier is based on the shape and distance similarities between the training samples and the edge fragments. Then the Cascaded Adaboost is used to select discriminative edge fragments to form a series of star models going from simpler models to more complex models to detect objects. The experimental results show that the method is very robust to large variances in color, texture, and shape. The system has a 90% detection rate on a database of up-right ear dogs.

Object Detection Using Edge Fragment-Based Features