Pedestrian Detection by Probabilistic Component Assembly

We present a novel pedestrian detection system based on probabilistic component assembly. A part-based model is proposed which uses three parts consisting of head-shoulder, torso and legs of a pedestrian. Components are detected using histograms of oriented gradients and Support Vector Machines (SVM). Optimal features are selected from a large feature pool by boosting techniques, in order to calculate a compact representation suitable for SVM. A Bayesian approach is used for the component grouping, consisting of an appearance model and a spatial model. The probabilistic grouping integrates the results, scale and position of the components. To distinguish both classes, pedestrian and non-pedestrian, a spatial model is trained for each class. Below miss rates of 8% our approach outperforms state of the art detectors. Above, performance is similar.