特征增强损失与前景注意力人群计数网络

Crowd Counting Network Based on Feature Enhancement Loss and Foreground Attention

人群计数旨在准确估计图像中的总人数并呈现其分布。相关数据集中的图像通常涉及各类场景且包含多人。为节约人力,大多数数据集通常在每个人头部以单点标注作为标签。然而,点标签无法囊括人头部的完整范围,使得人群特征与分布标签的匹配难以收敛,预测值无法聚集在前景区域,严重影响密度估计图质量和模型计数准确度。为了解决这个问题,使用计数损失来约束全图上的预测值范围,并佐以像素级的分布一致损失优化密度图匹配过程。此外,复杂场景中存在许多易与人群特征混淆的背景噪声,为了避免假阳性预测对后续计数和密度图估计的干扰,提出前景分割模块和特征增强损失来自适应地聚焦前景区域,并加大前景位置上人头特征对计数的贡献,从而达到抑制背景误判的作用。此外,为了使网络更好地适应人头的多尺度形态,对每个待训练图片分别进行上下采样操作,以获得具有同目标的多尺度形态。在多个数据集上进行了实验,结果表明,与最先进的方法相比,所提方法取得了更好或更有竞争力的结果。

Crowd counting aims to estimate the total number of people in an image and present its distribution accurately.The images in the relevant datasets usually involve a variety of scenes and include multiple people.To save labor,most datasets usually annotated each human head by a single point.However,the point labels cannot cover the full human head,which makes it difficult to converge the matching between the crowd feature and the distribution label,and the predicted values cannot be gathered in the foreground region,which seriously affects the density estimation map quality and count accuracy.To solve this problem,count loss is used to constrain the range of predictions on the full map,and a pixel-level distribution consistency loss is used to optimize the density map matching process.In addition,there are many background noises that are easily confused with crowd feature in complex scenes.In order to avoid the interference of false positive predictions on subsequent counting and density map estimation,a foreground segmentation module and feature enhancement loss are proposed to adaptively focus the foreground region and increase the contribution of human head features to the counts,so as to suppress background misjudgments.In addition,in order to make the network adapt to the multi-scale pattern of the human head better,up and down sampling operations are performed on each image to be trained to obtain the multi-scale pattern with the same object.Experiments on several datasets show that the proposed method achieves better or competitive results compared with state-of-the-art methods.