【摘要】 针对人体姿态估计任务特征提取精度低的问题，设计了一种特征高分辨率表征的姿态估计方法。该方法基于Mask RCNN模型，在RCNN获得目标实例后，将实例的检测结果映射到特征金字塔的高分辨率特征层，然后通过关键点预测模块上采样提高特征分辨率，最后将预测特征进行空间位置编码成分类任务，来获取每个实例的关键点二维空间位置并实现多人姿态估计。提出的多人姿态估计算法其检测任务性能精度提高1.0%,姿态估计任务精度提高0.5%,该算法减少了姿态估计的预测误差，进一步提升了姿态估计性能，提高了多人姿态估计的精度，实现了高分辨率表征。更多还原

【Abstract】 Aiming at the problem of low feature extraction accuracy of human pose estimation task, a pose estimation method based on high resolution feature representation was designed. This method is based on the Mask RCNN model. After obtaining the target instance in RCNN, the detection results of the instance are mapped to the high-resolution feature layer of the feature pyramid, and the feature resolution is improved through up-sampling of the key point prediction module. Finally, the spatial location of the predicted features is encoded into a classification task to obtain the key points of each instance two-dimensional space position and to achieve multi-person attitude estimation. The proposed multi-person attitude estimation algorithm based on the high-resolution representation improves the performance accuracy of the detection task by 1.0% and the accuracy of attitude estimation task by 0.5%. The algorithm reduces the prediction error of attitude estimation, and further improves the performance of attitude estimation with improved accuracy of multi-person attitude estimation and higher resolution representation.更多还原