基于改进自顶向下的行人运动预测方法  

A Pedestrian Motion Prediction Method Based on Improved Two-step Framework

作　　者：王淑青[1] 刘逸凡 WANG Shuqing;LIU Yifan(School of Electrical and Electronic Engin.,Hubei Univ.of Tech.,Wuhan 430068,China)

机构地区：[1]湖北工业大学电气与电子工程学院,湖北武汉430068

出　　处：《湖北工业大学学报》2020年第4期6-10,共5页Journal of Hubei University of Technology

基　　金：国家自然科学基金青年基金项目(61603127)。

摘　　要：针对传统的在复杂环境中多人姿态检测的自顶向下算法存在依赖人体框检测而导致的单人姿态检测错误的现象,提出一种基于改进的自顶向下的多人姿态检测并应用于行人运动预测的方法。该算法通过处理空间变换网络、单人姿势检测、反空间变换网络处理后的图像提取人体的骨点和姿势,经光流处理和长短期记忆神经网络的训练,预测行人接下来的动作。所述算法与6种经典多人姿态检测算法对比分析,实验结果表明:该方法得到的多人姿态检测图像准确,无冗杂人体框,无冗杂骨点,无骨骼交叉情况,行人运动预测效果良好。Aiming at the problem that traditional two-step framework algorithm of multi-person posture detection in complex environment relies on human frame detection,this paper proposes an improved two-step framework algorithm of multi-person posture detection method and applies it to pedestrian motion prediction.The algorithm extracts human bones and postures from images processed by spatial transformation network,single-person posture detection and anti-spatial transformation network,and predicts pedestrian’s next actions through optical flow processing and training of long short-term memory network.Compared with six classical multi-person posture detection algorithms,the experimental results show that the multi-person posture detection image obtained by this method is accurate,without redundant human frame,no redundant bone points,no skeleton crossover,and the pedestrian motion prediction effect behaves well.

关 键 词：自顶向下法 空间变换网络 单人姿势检测 光流处理 

分 类 号：TP242.2[自动化与计算机技术—检测技术与自动化装置]