MAPD: An improved multi-attribute pedestrian detection in a crowd

Abstract

Recently, CNN (convolutional neural networks) based pedestrian detection has made significant progress, but pedestrian detection in a crowd is still a challenge. Pedestrians are close to each other in a crowd, which is difficult for discriminating individuals. To address this issue, we propose an improved Multi-attribute pedestrian detection (MAPD) method, which coptimize person intra-class compactness and inter-class discrepancy. The contributions are fourfold: (1) We analyze the effect of positive setting on the detector and adopt a better positive settings strategy to mitigate extreme class imbalance problems. (2) Inspired by Person Reid, we employ the triplet loss function to learn the advanced id feature of pedestrians. (3) We propose a novel Piecewise NMS algorithm to reduce false positive of small objects. (4) We propose a novel multi-attribute NMS algorithm based on Piecewise NMS algorithm and id information, which can adaptively distinguish predicted boxes of different pedestrians and improve the detector performance. Finally, we evaluate the MAPD detector on two benchmark datasets, including CityPersons and CrowdHuman. Results show that our approach outperforms state-of-the-art methods with a big margin.