Abstract:

Deep learning has the advantages of autonomous learning object features, high recognition rate and strong robustness. Current human detection methods based on deep learning cannot adapt to changing scale of the target effectively. To solve this problem, this paper proposes a multi-scale and multi-human detection method, by combining the FPN feature pyramid with the two stages of the Faster R-CNN network respectively. At the same time, the number of positive and negative anchor points generated in RPN phase is balanced, and a more suitable aspect ratio of anchor points is adopted to optimize the original network. The results of experiments based on standard datasets PETS 2009, Caltech and INRIA show that, the proposed detection method performs better than the mainstream deep learning object detection algorithm.

Key words: deep learning, multi-scale and multi-human detection, Faster R-CNN network, FPN network, RPN network, anchor

摘要：

深度学习具有自主学习目标特征、识别率高、鲁棒性强等优点，当前基于深度学习的人体目标检测方法不能有效地适应目标的尺度变化。针对上述问题，提出多尺度多人的目标检测方法，将FPN特征金字塔分别与Faster R-CNN网络的两个阶段结合，同时，平衡RPN阶段产生的正负锚点的数量比例，并采用了更适合的锚点纵横比，对原始网络进行了一系列的优化。在标准数据集PETS 2009、Caltech和INRIA上的实验结果表明，提出的检测方法性能优于主流深度学习目标检测算法。

关键词: 深度学习, 多尺度多人目标检测, Faster R-CNN网络, FPN网络, RPN网络, 锚点