Efficient Pedestrian Detection in the Low Resolution via Sparse Representation with Sparse Support Regression.

We propose a novel pedestrian detection approach in the extreme Low-Resolution (LR) images via sparse representation. Pedestrian detection in the extreme LR images is very important for some specific applications such as abnormal event detection and video forensics from surveillance videos. Although the pedestrian detection in High-Resolution (HR) images has achieved remarkable progress, it is still a challenging task in the LR images, because the discriminative information in the HR images usually disappear in the LR ones. It makes the precision of the detectors in the LR images decrease by a large margin. Most of the traditional methods enlarge the LR image by the linear interpolation methods. However, it can not preserve the high frequency information very well, which is very important for the detectors. For solving this problem, we reconstruct the LR image in the high resolution by sparse representation. In our model, the LR and HR dictionaries are established respectively in the training stage, and the representative coefficients mapping relations are determined. Moreover, for improving the speed of feature extraction, the feature reconstruction in the LR images is converted to the sparse linear combination between the coefficients and the response of the atoms in HR dictionary by the LR-HR mapping, no matter how complex the feature extraction is. Experiments on the four challenging datasets: Caltech, INRIA, ETH and TUD-Brussels, demonstrate that our proposed method outperforms the state-of-the-art approaches and is much efficient with more than 10 times speedup.