Enhanced Single Shot MultiBox Detector for Pedestrian Detection
Yongren Cheng, Changhong Chen, Zongliang Gan · Proceedings of the 3rd International Conference on Computer Science and Application Engineering · 2019
Our work focus on pedestrian detection with the enhanced Single Shot MultiBox Detector (SSD) [1]. We find that the most of the missed pedestrian targets are concentrated on scenes with dense targets and small-scale objects when we regard SSD as the detector. In order to promote the detection performance for small-scale pedestrian targets, an additional feature in the third convolutional block and dense connections among convolutional blocks are added to the original SSD network structure. In addition, our study found that fine-grained sampling of the global prior box could improve the detection accuracy for multiscale pedestrians. We change the matching strategy between global prior boxes and truth target boxes, and add an extra loss that can keep more information in dense targets situation. The experimental results on PRW [2], Caltech and VOC dataset demonstrate that our enhanced SSD detector can achieve competitive detection accuracy as well as real-time detection speed on pedestrian objects.