Visible Feature Guidance for Crowd Pedestrian Detection

Abstract

Heavy occlusion and dense gathering in crowd scene make pedestrian detection become a challenging problem, because it's difficult to guess a precise full bounding box according to the invisible human part. To crack this nut, we propose a mechanism called Visible Feature Guidance (VFG) for both training and inference. During training, we adopt visible feature to regress the simultaneous outputs of visible bounding box and full bounding box. Then we perform NMS only on visible bounding boxes to achieve the best fitting full box in inference. This manner can alleviate the incapable influence brought by NMS in crowd scene and make full bounding box more precisely. Furthermore, in order to ease feature association in the post application process, such as pedestrian tracking, we apply Hungarian algorithm to associate parts for a human instance. Our proposed method can stably bring about 2∼3% improvements in mAP and AP50 for both two-stage and one-stage detector. It's also more effective for MR −2 especially with the stricter IoU. Experiments on Crowdhuman, Cityperson, Caltech and KITTI datasets show that visible feature guidance can help detector achieve promisingly better performances. Moreover, parts association produces a strong benchmark on Crowdhuman for the vision community.

Fast and accurate vision-based pedestrian detection

“…1.7b) of pedestrian body part. The visibility information is then exploited to guide feature extraction for pedestrian [13,27,[46][47][48]. This strategy alleviates the problem of feature misalignment for occluded pedestrians but requires visible part information during training, which increases annotation costs and limits their applications in real-world scenarios.…”

“…A novel attribute feature-based NMS is designed to distinguish a person from a highly occluded group by adaptively rejecting the false detections in crowded scenes. FRCNN+VFG [47] proposes to simultaneously conduct full body and visible body prediction, and then performs NMS only on visible body prediction. R 2 NMS [163] proposes a representative region NMS to exploit visible pedestrian parts to effectively discard the redundant boxes without introducing many false detections, in which the visible body and full body are predicted as a pair to ensure a strong correlation between the two predicted boxes.…”

“…used in [13,27,47,48,157,158,160,163]. The increased cost of annotations limits their application in real-world scenarios.…”