Object Detection Using a Cascade of Classifiers

Typical object detection systems work by training a classifier on features extracted at different scales of an object. In this paper we investigate the performance of an object detection system in which different classifiers which are trained at various scales of an object are combined and compare the performance with a typical object detection system where a single classifier is trained for all the scales. The notion behind such an approach is that the features extracted over smaller scales give more objectpsilas specific information whereas large scale features provide more contextual information. We trained different classifiers for different scales and combined their output to reach a decision about the existence of an object. Confidence rated Ada-boost is used to train the classifiers. It was found that training a single classifier for all the scales results in superior performance as compared to training different classifiers for each scale and than combining their results. We show our results on objects belonging to three categories in TUDarmstadt and one category in Caltech4.