Pedestrian Detection: Exploring Virtual Worlds

State-of-the-art pedestrian detectors rely on machine learning algorithms trained with labelled samples, i.e., examples (pedestrians), and counterexamples (background). Therefore, in order to build robust pedestrian detectors the quality of the training data is fundamental. In this chapter, we explore the possibilities that a virtual computer generated database, free of real-world images, can offer to this process. We record training sequences in realistic virtual cities and train appearance-based pedestrian classifiers using HOG and linear SVM, a baseline method for building such classifiers that remains competitive for pedestrian detection. We test such classifiers on publicly available real-world datasets. Besides, we present specific analysis on the required number of virtual models and training examples to get a satisfactory performance, and present results on how the pose influences the performance.