Selecting and evaluating data for training a pedestrian detector for crowded conditions

Abstract

Computer vision algorithms for pedestrian detection are often based on classification derived from supervised learning and therefore require training data, which can be built by using generic or specific images. In this field, INRIA datasets are a standard reference but include only few CCTV camera samples. Therefore, for a CCTV camera system it might be interesting to have specific training data. However, in practice it is impossible to create a training data for each camera view. Thus, this paper presents an evaluation of a pedestrian detection algorithm in crowded conditions in relation to the training data, and shows that a CCTV camera training data provides better results and can be reused for similar CCTV camera views.