Looking at People - Pedestrian Classification The ability to recognize humans and their activities by vision is key for a machine to interact intelligently and effortlessly with a human-inhabited environment. This MS Thesis deals with detecting people in images (i.e. pedestrians), which are located at a certain distance to the camera. We aim to address the problem by a pattern classification approach. Rather than trying to locate various body parts (e.g. head, hands, feet) in images explicitly based on prior knowledge about human appearance, we describe a region of interest in terms of low-level features and aggregate the latter into a feature vector. In an off-line training phase, a pattern classifier derives an internal pedestrian representation using a large numbers of previously categorized feature vectors. In the on-line recall phase the derived representation is used to classify unknown samples. Issues to be resolved are data normalization (e.g. size, contrast), feature selection (e.g. normalized pixel intensities, wavelets), dimensionality reduction (e.g. PCA, ICA) and actual pattern classification (e.g. SVM, NN). Extensive image data, both from the visible and infrared spectrum, is available from DaimlerChrysler. Existing MATLAB toolboxes for pattern classification can be used. System implementation for the recall phase is under C/C++, with emphasis on real-time considerations. The MS Thesis may lead to a publication.
Keywords:	Computer Vision, Looking at People, Pedestrian Recognition
Study:
Contact:	Dariu Gavrila
Status:	open
Location:	University of Amsterdam
References:	Looking-at-People survey , Pedestrian Detection survey , Gavrila website
Talks: