基于改进FCOS的拥挤行人检测算法

针对大规模拥挤场景视频中行人目标小、行人遮挡和行人交叠而导致的检测困难等问题，本文将逐像素预测目标检测框架—全卷积单阶段目标检测FCOS（fully convolutional one-stage object detection）应用于行人检测，提出一种改进的主干网络用于提取行人特征，通过增加尺度回归实现目标行人的多尺度检测，同时减少其他特征层检测的目标数量，进而提升行人检测的能力。在拥挤行人场景数据集CrowdHuman和小目标行人数据集Caltech上的大量实验结果表明，和目前先进的方法相比，本文的方法对行人的检测精度有所提升，特别是对于小目标行人检测。与原始FCOS算法相比，在CrowdHuman上平均精度提升接近15%，丢失率降低接近33.0%；在Caltech上的平均精度提升2%。在复杂拥挤场景下的实际应用也证明本文方法的有效性。

Crowded pedestrian detection algorithm based on improved FCOS

In view of the detection difficulty resulting from small pedestrian objects, pedestrian occlusion, and pedestrian overlap in large-scale crowded scene videos, this study applies a pixel-by-pixel prediction object detection framework, i.e., fully convolutional one-stage object detection (FCOS), for pedestrian detection. An improved backbone network is proposed to extract pedestrian features, achieve multi-scale detection of object pedestrians by increasing scale regression, reduce the number of objects detected by other feature layers, and thereby improve the ability of pedestrian detection. Several experiments have been performed on the crowded pedestrian scene dataset CrowdHuman and the small object pedestrian dataset Caltech. The results show that compared with current advanced methods, the proposed algorithm makes some improvements in the pedestrian detection accuracy, especially for small object pedestrian detection. Compared with the original FCOS framework, the average precision on CrowdHuman is increased by nearly 15% and the miss rate is decreased by nearly 33.0%. The average precision on Caltech is increased by 2%. Moreover, the actual use in complex, crowded scenarios proves the effectiveness of this algorithm.