Representative Template Set Generation Method for Pedestrian Detection

Abstract

Template matching is an effective approach for pedestrian detection. In order to achieve real-time and accurate detection, how to obtain a suitable representative template set is still an open problem due to the large variety of pedestrian shape. This paper introduced a representative template generation method for a template matching based pedestrian detection system (PDS). Based on nonlinear manifold learning and clustering, the new approach can generate a suitable representative template subset from a large amount of original templates. First, an improved nonlinear dimensionality reduction method was proposed to map original templates to feature vectors (points) in the low-dimensional embedding space; second, representative points were generated in the embedding space by clustering; at the end, corresponding representative template set were synthesized by mapping inversely the newly generated points from the embedding space to the visual input space.The experimental results showed that the template generation method speeds up detection procedure without considerable loss of performance.