Pyramidal Statistics of Oriented Filtering for robust pedestrian detection

We study the problem of robust pedestrian detection. A new descriptor, Pyramidal Statistics of Oriented Filtering (PSOF), is proposed for shape representation. Unlike one-scale gradient-based methods, the PSOF descriptor constructs an image pyramid and uses a Gabor filter bank to obtain multi-scale pixel-level orientation information. Then, locally normalized pyramidal statistics of these Gabor responses are used to represent object shape. After feature extraction, the AdaBoost training algorithm is adopted to train a classifier for the final pedestrian detector. We show experimentally that the PSOF descriptor is much more robust to image blur and noise than the HOG (Histograms of Oriented Gradients) descriptor, as well as possesses excellent detection performance in normal imaging condition as HOG does. We also study the influence of various parameter settings, concluding that multi-scale information and statistic combination are two important factors for the robustness of the PSOF descriptor.