摘要
实时、高精度的行人三维姿态估计可有效提升无人驾驶汽车智能化水平。在Open Pose深度学习基础上,提取人体2D关节点,通过建立简化人体躯干数学模型来计算躯干空间位置和姿态,并以单个肢体为对象,通过分层式优化计算求解人体四肢关节角,进而确定人体3D姿态。主、客观实验测试结果表明,行人各关节空间位置平均误差低于78 mm,算法平均处理时间每帧低于52 ms。
Real-time,high-precision pedestrian 3D pose estimation can improve the intelligent level of driverless vehicles effectively.Based on the OpenPose deep learning,the 2D key points of the human body was extracted,the human trunk position and pose were calculated through establishing a simplified human torso math model.And on a single limb,the joint angle of the human limbs through hierarchical optimization were calculated,so as human body 3D pose were gotten.The subjective and objective experiments show that the average error of joint space position is less than 78 mm,and the detection time of each frame is less than 52 ms.
作者
徐彬
郑燕萍
曹高兴
XU Bin;ZHENG Yan-ping;CAO Gao-xing(Northern Navigation Technology Group Co., Ltd.3, Beijing 100070, China;School of Automotive and Transportation Engineering,Nanjing Forestry University, Nanjing 210037, China;Northern Navigation Technology Group Co., Ltd.3, Beijing 100070, China)
出处
《科学技术与工程》
北大核心
2018年第34期85-91,共7页
Science Technology and Engineering
基金
2017年江苏省重点研发计划(BE2017008)资助.
关键词
无人驾驶
行人姿态估计
多点透视
分层优化
driverless vehicles
pedestrian pose estimation
multi-points fluoroscopy
hierarchical optimization
