目的 以有标记点三维运动捕捉系统（MoCap）为金标准，基于双向长短时记忆（bi-lateral long short term memory，BiLSTM）递归神经网络和线性回归算法构建深度学习融合模型，减小深度传感器的系统误差，从而提高深度传感器下肢运动学分析的准确性。方法 招募10名健康男性大学生进行步态分析，应用MoCap系统和Kinect V2传感器同时采集数据。通过Cleveland Clinic及Kinect逆运动学模型分别计算下肢关节角度。以MoCap系统为目标，Kinect系统得到的角度为输入构建数据集，分别用BiLSTM算法和线性回归算法构建学习模型，得到系统误差修正后的下肢关节角度。使用留一交叉验证法评估模型的性能。采用多重相关系数（coefficient of multiple correlations，CMC）及均方根误差（root mean square error，RMSE）表示下肢关节角度波形曲线相似程度以及平均误差。结果 BiLSTM网络比线性回归算法更能够处理高度非线性的回归问题，尤其是在髋关节内收/外展、髋关节内旋/外旋和踝关节趾屈/背屈角度上。应用BiLSTM网络的误差修正算法显著降低Kinect的系统误差（RMSE<10°，其中髋关节RMSE<5°），下肢角度波形呈现很好的一致性（除髋关节内旋/外旋角度外，CMC>0.7）。结论 本文开发的基于深度学习融合模型的无标记点步态分析系统可以准确评估下肢运动学参数、关节活动能力、步行功能等，在临床和家庭康复中具有广泛的应用前景。
Objective Taking three-dimensional (3D) motion capture system (MoCap) as the gold standard, a deep learning fusion model based on bi-lateral long short-term memory (BiLSTM) recurrent neural network and linear regression algorithm was developed to reduce system error of the Kinect sensor in lower limb kinematics measurement. Methods Ten healthy male college students were recruited for gait analysis. The 3D coordinates of the reflective markers and the lower limb joint centers were simultaneously collected using the MoCap system and the Kinect V2 sensor, respectively. The joint angles of lower limbs were calculated using the Cleveland clinic kinematic model and the Kinect kinematic model, respectively. The dataset was constructed using the MoCap system as the target and the angles via the Kinect system as the input. A BiLSTM network and a linear regression model for all lower limb angles were developed to obtain the refined angles. A leave-one subject-out cross-validation method was employed to study the performance of the models. The coefficient of multiple correlations (CMC) and root mean square error (RMSE) were used to investigate the similarity and the mean deviation between the joint angle waveforms via the MoCap and the Kinect system. ResultsIn comparison with the linear regression algorithm, the BiLSTM had better performance in the aspect of dealing highly nonlinear regression problems, especially for hip flexion/extension, hip adduction/abduction, and ankle dorsi/plantar flexion angles. The deep learning refined model significantly reduced the system error of Kinect. The mean RMSEs for all joint angles were mainly smaller than 10°, and the RMSEs of the hip joint were smaller than 5°. The joint angle waveforms presented very good similarity with the golden standard. The CMCs of joint angles were greater than 0.7 except for hip rotation angle. Conclusions The markerless gait analysis system based on deep learning fusion model developed in this study can accurately assess lower limb kinematics, joint mobility, walking functions, and has good prospect to be applied in clinical and home rehabilitation.