Deep Pedestrian Detection Using Contextual Information and Multi-level Features

Recently, Faster R-CNN achieves great performance in deep learning based object detection. However, a major bottleneck of Faster R-CNN lies on the sharp performance deterioration when detecting objects that are small in size or have a similar appearance with their backgrounds. To address this problem, we present a new pedestrian detection approach based on Faster R-CNN, which combines contextual information with multi-level features. The contextual information is embedded by pooling information from a larger area around the original region of interest. It helps pedestrians detection from cluttered backgrounds. The multi-level features can be obtained by pooling proposal-specific features from several shallow but high-resolution layers. These features are more informative for detecting small-size pedestrians. Extensive experiments on the challenging Caltech dataset validate that our approach not only performs better than the baseline of Faster R-CNN but also boosts the detection performance when combined with contextual information and multi-level features. Meanwhile, compared with numerous pedestrian detection approaches, our combined method outperforms all of them and achieves a quite superior performance.