A Multi-Feature Part-Based Object Detection System

Google Tech Talks May 25, 2007 - Speaker: Bernd Heisele, Honda

ABSTRACT
I will start with an overview on object recognition systems which use local features and ... all » analyze their strengths and weaknesses. I will then present a general purpose part-based object detection system which we evaluated on a benchmark pedestrian detection data set . In a first step, the system computes feature maps from the training images. It then randomly extracts a large number of rectangular parts from the feature maps and clusters the parts based on their feature similarity and their x-y-location in the feature maps. The cluster centers build an initial set of part templates from which the system selects a subset using the gentle-boost algorithm. The localization of the parts during classification is performed by normalized cross-correlation of the part templates with feature maps. Three different types of feature maps were used in our experiments: Original gray value images, the magnitudes of the gradient, and Gabor filtered images. In experiments on a benchmark pedestrian detection database, we investigate how the number of the components, the feature type and the training data affects the detection performance. The system is compared to state-of-the-art pedestrian detectors.