物理约束与神经网络混合驱动的水黾姿态估计方法  

Water Strider Pose Estimation Method Driven by Hybrid of Physical Constraints and Neural Networks

作　　者：叶瑛歆 黄聪 朱文强 Ye Yingxin;Huang Cong;Zhu Wenqiang(School of Mechanical and Electrical Engineering,Shandong Jianzhu University,Jinan 250101,China;Jinan Gold Phoenix Brake Systems Co.,Ltd.,Jinan 251499,China)

机构地区：[1]山东建筑大学机电工程学院,济南250101 [2]济南金麒麟刹车系统有限公司,济南251499

出　　处：《机电工程技术》2025年第3期114-120,共7页Mechanical & Electrical Engineering Technology

基　　金：山东省自然科学基金青年项目(ZR2021QE128);山东省重点研发计划项目(2022CXGC010101)。

摘　　要：准确的动物姿态估计是进行动物运动特性分析的关键。由于水黾运动主要依靠中腿,并且在水面运动时中腿介于水-空气中,进行姿态估计难度大,并且存在中腿关节不好分辨的情况。为了提高复杂环境下水黾姿态估计的准确性,提出了一种物理约束与神经网络混合驱动的水黾运动姿态估计方法:利用DeepLabCut神经网络对水黾中腿各关节在图片中的像素坐标进行预测,然后将水黾中腿部分关节的特性作为物理约束,利用双目相机能够获取深度及距离信息的特性,对神经网络预估计的坐标进行修正,得到修正后新的水黾姿态估计信息,结合双目标定矩阵信息将左右相机中的二维姿态估计信息变换到三维中,实现水黾的三维姿态估计。实验结果表明介绍的水黾中腿姿态估计方法实现了89.11%的高准确率,相较于仅依赖神经网络训练的识别预测,准确率提升了约10.3%。实现了较高精度的水黾姿态估计,并且物理约束与神经网络混合驱动的方法比单独使用神经网络预测方法的精度高。Accurate animal pose estimation serves as a crucial foundation for analysis of animal locomotory characteristics.Given that water striders primarily rely on their middle legs for locomotion,with these legs straddling the water-air interface during surface movements,pose estimation poses significant challenges,particularly in differentiating middle leg joints.To improve the accuracy of pose estimation under complex environments,a hybrid approach combining physical constraints and neural networks for water strider motion pose estimation is proposed.Specifically,DeepLabCut neural network is firstly utilized to predict the pixel coordinates of each joint in the water strider’s middle legs within images.Subsequently,physical constraints are incorporated based on the unique characteristics of the middle leg joints,leveraging the binocular camera’s capability to acquire depth and distance information to refine the pre-estimated coordinates from the neural network,yielding refined pose estimation data.By integrating the stereo calibration matrix,the two-dimensional pose estimates are transformed from both cameras into a three-dimensional space,achieving three-dimensional pose estimation of the water strider.Experimental validation in a tailored setup acquires a high precision up to 89.11%,which is about 10.3%higher than standalone neural network predictions,demonstrates the achievement of high-precision pose estimation,with the hybrid method of physical constraints and neural networks outperforming standalone neural network predictions in terms of accuracy.

关 键 词：仿生机器人 姿态估计 神经网络 混合驱动 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]