Steerable second order intensity features for pedestrian detection

In this paper we consider how second order intensity features can be used to boost the accuracies of SVM-based pedestrian detectors, which remain much faster than recently popular deep learning approaches. In particular, we demonstrate that combining second order information features, corresponding to Hessians of patch-wise image intensities, with HOG and LBP features leads to more than 10% improvement in accuracy for frequently used and difficult datasets. In addition, we present a framework to visualize the responses of the linear SVM classifier at different locations and for different feature types, which enables a comprehensive and detailed analysis of failure modes of the current HOG-LBP detector. An interesting observation made is that the weight patterns of the Hessian-based features change when combined with HOG and LBP features as compared to using Hessians only, and that these changes can be understood from an intuitive perspective as detailed in our paper.