Part-based pedestrian detection using grammar model and ABM-HoG features

To handle the pedestrian appearance and pose variations in complex traffic environments, we present one part-based pedestrian detection approach using a stochastic grammar model in this paper. The And-Or graph model is introduced to represent the human body as an assembly of compositional and reconfigurable parts. Thus, the task of detection is converted into the human parsing problem, which is a Bayesian inference process. We model the appearance of pedestrian parts in a rich feature representation. This appearance model enhances the Histogram of Gradients (HoG) map with Active Basis Model (ABM), which is a sparse deformable template depicting salient structures of objects. Then, geometry constraints among parts are described by Gaussian distributions. Finally, the bottom-up parsing inference is conducted by aggregating scores to get the pedestrian detection responses. In experiments, we show the superiority of our appearance model, as well as the reliable pedestrian detection results of our approach in complex traffic scenes.