基于PRCNN-Attention鲁棒零速检测的行人惯性导航方法

Pedestrian inertial navigation method based on PRCNN-Attention robust zero-velocity detection

针对零速修正的行人惯性导航方法中MEMS IMU误差大且复杂多变、复杂运动类型下传统阈值零速检测鲁棒性差的问题,提出一种基于PRCNN-Attention鲁棒零速检测的导航方法。首先,引入注意力机制,突出不同运动类型下的关键信息,设计了基于并联式循环卷积网络的深度神经网络(PRCNN-Attention)框架,实现对零速状态的鲁棒识别;然后,将零速信息作为量测信息,基于不变扩展卡尔曼滤波器实现信息融合。最后,在公开数据集以及实际场景中对所提方法进行了验证。实验结果表明,相较于固定阈值零速检测方法和自适应阈值零速检测方法,所提方法在公开数据集上的零速检测精确率分别由0.752和0.920提升到0.978,导航误差分别降低了54.1%和31.8%,在实际场景中导航误差分别降低了28.1%和13.5%,验证了所提方法的鲁棒性和有效性。

Aiming at the problems of large MEMS IMU errors,complex and variable,and poor robustness of traditional threshold zero-speed detection under complex motion types in the pedestrian inertial navigation method based on zero-velocity update,a pedestrian inertial navigation method based on PRCNN-Attention robust zero-speed detection is proposed.Firstly,the attention is introduced to highlight the key information under different types of motion,and a deep neural network(PRCNN-Attention)framework based on parallel recurrent convolutional network is designed to realize the robust recognition of zero-velocity state.Then,zero-velocity information is used as the measurement information,and the information fusion is realized based on an invariant extended Kalman filter.Finally,the proposed method is verified in the public data set and the actual scene.The experimental results show that compared with the fixed threshold zero velocity detection method and the adaptive threshold zero velocity detection method,zero velocity detection accuracy of the proposed method on the public dataset is increased from 0.752 and 0.920 to 0.978,and the navigation error is reduced by 54.1%and 31.8%.The navigation error in the actual scene is reduced by 28.1%and 13.5%,respectively,which verifies the robustness and effectiveness of the proposed method.