基于改进CPN的人体关键点定位算法研究

Human Keypoint Detection Algorithm Based on Improved CPN

人体骨骼关键点定位是计算机理解图片或视频中人物行为的重要步骤,但定位的准确率很容易受到遮挡、干扰、复杂背景等因素的影响。把注意力模型引入CPN特征金字塔网络对之进行改进,以提升网络定位的准确率。注意力模型CBAM通过学习权重,分别对特征图通道和空间像素进行加权,实现自动感知特征图各区域的重要程度。金字塔网络通过上采样等方式将深层和浅层特征融合起来。该方法利用COCO数据集进行测试和对比实验,结果表明,该方法改善原CPN网络及当前方法在遮挡、干扰、背景复杂等情况下检测不准的问题,平均准确率和平均召回率均有所提高。与当前主要关键点定位算法相比,该算法的定位准确率或精度有一定优势。

Human keypoint detection is an important step for computers to understand the behavior of people in pictures or videos,but its accuracy is easily affected by surrounding occlusion,interference and complicated background.In this paper,attention model is introduced into Cas⁃caded Pyramid Network to improve the network accuracy.By learning coefficients,the attention model CBAM weights the channels and spa⁃tial pixels of the feature map,so as to automatically perceive the importance of each region of the feature map.The pyramid network inte⁃grates the deep and shallow features by means of up-sampling.The method is tested and evaluated on COCO dataset,of which the results show that it improves the low accuracy of the original CPN network and the current methods in the case of occlusion,interference,and com⁃plicated background.The average accuracy and recall have been improved.And compared with the current keypoint detection algorithm,the method in this paper has its certain advantages.