基于改进堆叠沙漏网络的交警姿态估计

Traffic police pose estimation based on improved stacked hourglass networks

针对交警动作中的姿态估计问题,提出一种改进的堆叠沙漏网络模型。该模型通过减少沙漏网络级联次数,来简化堆叠沙漏网络结构。利用多尺度下深浅层特征信息之间的聚合,得到丰富的上下文信息,增强姿态、遮挡、低分辨率图像的鲁棒性。将不同阶段产生的热图估计结果进行融合平均化处理,进一步提高局部位置坐标的精细定位以及整体估计结果的准确性。在MPII数据集以及中国交警数据集上进行实验,结果表明,改进后的网络模型提高了运行的效率,同时可以很好地对目标交警的姿态特征信息进行提取,对结果热图平均化处理后,提高了位置坐标整体估计的准确性。

An improved stacked hourglass network model is proposed for the problem of traffic police pose estimation.The model simplifies the stack hourglass network structure by reducing the number of cascades.The aggregation of information from multi-scale deep shallow features is used to obtain rich contextual information and enhance the robustness of the image with pose,hiding and low resolution.The heat map estimation results generated at different stages were averaged by fusion to further improve the precision positioning of local coordinates and the accuracy of overall estimation results.Experiments on the MPII data set and China Traffic police data set show that the improved network model improves the operation efficiency,and can extract the attitude characteristic information of the target traffic police well.After averaging the result heat map,the accuracy of the overall position coordinate estimation is improved.