Integration Network for Fast Pedestrian Detection in Crowd Scenarios

Pedestrian detection has experienced tremendous improvement in recent years due to advanced detector structure based on deep learning technique. However, occlusion caused by the crowd still remains one of challenges in pedestrian detection, since in many real-world scenarios pedestrians tend to gather together thus cause serious crowd occlusion. To overcome this issue, we introduce a fast deep pedestrian detector called pedestrian detection with integrated segmentation context(PDIS) which benefits affluent semantic information by a segmentation feature extractor and integration branch. The detection branch and segmentation branch are based on fully convolutional network, which possesses numerous advantages in both tasks. Our detector and extractor are generated from detection branch and segmentation branch respectively which are supervised on separately ground-truth, bringing extensive useful information. The integration branch consists of several convolutional layers to fuse context of different dimensions. Experiments on the CrowdHuman dataset demonstrate that our approach maintains high detection performance at a fast speed.