Implementation and optimization of wavelet and symmetry features for vision-based pedestrian detection

This paper presents two of the most important parts of a vision-based pedestrian detection system: the feature extraction and the classification module. Wavelet-based features and a combination of symmetry and edge density features are extracted from a monochrome image captured by a vehicle-mounted camera and fed into an SVM-classifier, more precisely a modified version of libSVM [1]. For both types of features an optimization approach based on image masks is proposed. In order to weight the impact of classifier results (false negatives are preferred over more false negatives in the case of pedestrian detection) the F-measure is used as statistical measure. An overview on the advantages and drawbacks of the implemented features and the optimization approach is given, based on the results received from tests using pedestrian and non-pedestrian images extracted from video sequences showing urban traffic scenes.