Abstract:In crowded pedestrian detection, occlusion situations are common challenges that seriously impact detection performance. These occlusions are usually classified into pedestrian-to-pedestrian occlusions and object-to-pedestrian occlusions which result in false detection and missed detection. In this paper, we propose a novel model to address the crowded pedestrian detections in the cases of occlusions, which can generate an optimal bounding box containing the pedestrian instance with accurate position information. Firstly, Distance-Intersection over Union loss is introduced in Region Proposal Networks module for network training to generate proposal boxes, considering both the position and area of the region where the pedestrian is occluded. Secondly, a refinement module is added in Region Convolutional Neural Network to eliminate false positive proposal boxes, Earth Mover's Distance Loss is used to re-predict the pedestrian in these boxes. Finally, Relocation Non-Maximum Suppression is employed to select the optimal bounding box. Considering the parts of the pedestrian contained by its adjacent proposal boxes, the optimal bounding box is located in order to achieve the complete pedestrian instance. The proposed model is evaluated on CrowdHuman and CityPersons datasets respectively. On CrowdHuman dataset, the proposed model improves AP by 5.6% and JI by 5.2%, while reducing MR −2 by 3.8% compared to the baseline. Compared to the state-of-the-art model, the proposed model reduces 0.4% on MR −2 , which shows its effectiveness for pedestrian detection in crowded scenes. On CityPersons dataset, the proposed model obtains the AP with 96.8% among all the evaluated models, which indicates its generalization for pedestrian detections in various crowded scenes.

Improving Pedestrian Detection in Crowds with Synthetic Occlusion Images

Towards Accurate Dense Pedestrian Detection Via Occlusion-Prediction Aware Label Assignment and Hierarchical-Nms.

Pedestrian Detection Method Based on Improved YOLOv5s for Densely Occluded Scenarios

See Extensively While Focusing on the Core Area for Pedestrian Detection.

PedHunter: Occlusion Robust Pedestrian Detector in Crowded Scenes

Learning Pixel-Level and Instance-Level Context-Aware Features for Pedestrian Detection in Crowds.

Pedestrian Partial Occlusion Detection Method Based on Improved YOLO Network Structure

Attention-Guided Sample-Based Feature Enhancement Network for Crowded Pedestrian Detection Using Vision Sensors

Context feature fusion and enhanced non-maximum suppression for pedestrian detection in crowded scenes

Guided Attention in CNNs for Occluded Pedestrian Detection and Re-identification

Pushing the Limits of Deep CNNs for Pedestrian Detection

Crowded pedestrian detection with optimal bounding box relocation

Imagine the Unseen: Occluded Pedestrian Detection via Adversarial Feature Completion

Mask-Guided Attention Network for Occluded Pedestrian Detection

Survey of Pedestrian Detection with Occlusion

Pedestrian Detection in Crowded Crowd Scene Based on Deep Learning

A Boosted Multi-Task Model for Pedestrian Detection with Occlusion Handling.

Occluded Pedestrian Attention Network : an Occluded Pedestrian Detector

Improved YOLOv3 Integrating SENet and Optimized GIoU Loss for Occluded Pedestrian Detection

Attribute-Aware Pedestrian Detection in a Crowd

Occlusion and multi-scale pedestrian detection A review