拥挤场景下的行人姿态估计研究  

Research on Pedestrian Pose Estimation in Crowded Scenes

作　　者：任超[1] 赵波[1] 张伟伟 REN Chao;ZHAO Bo;ZHANG Weiwei(Shanghai University of Engineering Science,Shanghai 201620)

机构地区：[1]上海工程技术大学,上海201620

出　　处：《计算机与数字工程》2022年第6期1359-1364,共6页Computer & Digital Engineering

摘　　要：针对行人拥挤场景下姿态密集交叉、识别能力急剧下降等问题,论文提出了一种有效解决拥挤场景下多人检测以及姿态估计的方法。该方法主要由两个关键部分组成。首先利用Faster-RCNN行人检测器,预测出包含行人位置和大小规模的矩形框。然后,预测每个矩形框中可能包含的行人骨骼关键点,使用全卷积ResNet网络来预测每个骨骼关键点的热力图和偏移量并输出。使用基于骨骼关键点的非极大值抑制和热力图置信度分数预测的新方式,使得最终输出行人姿态结果更加准确,对于背影、遮挡等不可避免的干扰具有一定的鲁棒性。实验结果表明,该方法在COCO数据集和SUES行人数据集上获得平均精度达到了0.673,与之前在相同数据集上的同类方法相比,有不低于5%的改进。Aiming at the problems of crowded intersections and sharp decline in recognition ability in crowded pedestrian scenarios,this paper proposes an effective method to solve multi-person detection and pose estimation in crowded scenes.The method is mainly composed of two key parts.Firstly,the Faster-RCNN pedestrian detector is used to predict the rectangular box containing the pedestrian position and size.Then the pedestrian skeleton key points that may be included in each rectangular box are predicted,and the fully convolution ResNet network is used to predict each bone key point to output the heat map and offset.The use of a new method based on non-maximum suppression of bone key points and prediction of heat map confidence scores makes the final output pedestrian pose results more accurate,and it is robust to inevitable disturbances such as back view and occlusion.Experimental results show that this method achieves an average accuracy of 0.673 on the COCO data set and SUES pedestrian data set,which is not less than 5%improvement compared with the previous similar methods on the same data set.

关 键 词：行人检测 姿态估计 卷积神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]