基于语义分割注意力与可见区域预测的行人检测方法

为改善图像中遮挡和小尺寸行人的检测精度，提出一种基于语义分割注意力和可见区域预测的行人检测方法.具体地，在SSD(single shot multi-box detector)目标检测网络的基础上，首先优化SSD的超参数设置，使其更适于行人检测;然后在主干网络中引入基于语义分割的注意力分支来增强行人检测特征的表达能力;最后提出一种检测预测模块，它不仅能同时预测行人整体和可见区域，还能利用可见区域预测分支所学的特征去引导整体检测特征的学习，提升检测效果.在Caltech行人检测数据集上进行了实验，所提方法的对数平均缺失率为5.5%，与已有方法相比具有一定的优势.

Pedestrian Detection Based on Semantic Segmentation Attention and Visible Region Prediction

To improve the detection performance on occluded and small pedestrians in images， a pedestrian detection method based on semantic segmentation attention and visible region prediction was proposed. Specifically， based on the single shot multi-box detector(SSD)object detection network， the hyperparameter setting of the SSD was firstly optimized to make it more suitable for pedestrian detection. Then the semantic segmentation attention branch was introduced into the network to enhance the pedestrian detection features learned by the network. Finally， a detection prediction module which can simultaneously detect the full bodies and visible regions of pedestrians was developed. This module has the advantage of leveraging the features learned from visible regions to guide the learning of the full-body detection features， hence improving the overall detection accuracy. The experiment carried out on the Caltech pedestrian detection benchmark shows that the log-average miss rate of the proposed method is 5.5%， which is competitive compared with existing pedestrian detection approaches.