Generalizing Pedestrian Detector to Low Resolution Images by Integrating High and Low Resolution Features

Detecting low resolution (low-res) pedestrians is rather important for automotive driving systems as it can provide enough time for decision making. However, many detectors perform quite well on high resolution (high-res) pedestrians, while perform poorly on low resolution ones. In this paper, to generalize the high performance of high-res pedestrians detectors to low-res pedestrians ones, we propose to integrate the high and low resolution features, the resolution-aware transformations are learned simultaneously to map the high and low resolution pedestrians to a common space and the shared detector in this space can be learned based on the linear SVM model. Meanwhile, a Maximum Mean Discrepancy (MMD) is considered to minimize the distribution mismatch of the transformed high and low resolution features. Then the shared detector and the transformations are learned jointly by minimizing both the distribution mismatch between the mapped high and low resolution samples and structural risk functional. In addition, we show that our proposed learning framework can be converted to a standard Multiple Kernel Learning (MKL) problem, which is convex and hence the global solution can be guaranteed. Experiments on the Caltech Pedestrian Benchmark show that our method is effective for the low resolution pedestrian detection tasks.