【摘要】 在目前的行人检测方法中，中心尺度预测(center-scale prediction, CSP)模型具有检测速度快，无需预设锚点等优点。但是，CSP模型并没有针对行人遮挡问题提出解决方法。为此，在CSP模型的基础上，提出了一个基于可视注意力机制的中心尺度预测(visible attentionmechanism-basedCSP,VA-CSP)模型，同时预测行人及其可视区域的边界框，并构造一个中心-可视中心(center-visible center, C-V)变换预测分支，将行人及其可视区域匹配，使模型具有正确的可视注意力机制，提升遮挡行人的检测精度。在Citypersons和Caltech行人检测数据集上进行了实验，在Citypersons验证集的不同遮挡程度的子数据集Reasonable、Heavy、Partial和Bare上，得到了9.6%、48.1%、9.1%和6.6%的丢失率，相比CSP分别提升了1.4%、1.2%、1.3%和0.7%。在Caltech测试集的Reasonable子数据集上得到了3.2%的丢失率，相比CSP提升了1.3%。与其他目前最新的模型相比，所提模型具有更高的检测精度。更多还原

【Abstract】 Among existing pedestrian detectors, the center-scale prediction(CSP) model performs detection quickly and predicts the center and scale of objects directly without defining anchors. However, the detection accuracy of CSP is unsatisfactory due to the heavy occlusion and large preference variation in pedestrian detection. To handle this concern, we adopt a visible attention mechanism, which prompts the model to concentrate on the visible regions of pedestrians, which can alleviate the influence of occlusion and preference variation. A visible attention-based center-scale prediction model(VA-CSP) is proposed, which generates the center and scale of the pedestrian along with its visible region simultaneously. By constructing a center-visible center(C-V) transfer branch, the bounding box and visible bounding box of the same pedestrian are tightly associated. The constructed association can help our model capture the correct visible attention of the pedestrian. Comprehensive experiments are conducted on two challenging datasets, i.e., Caltech and Citypersons. Compared with other current models, the proposed model has higher detection accuracy.更多还原