Three-dimensional attitude estimation for unmanned aerial vehicles is usually based on the combination of magnetometer, accelerometer, and gyroscope (MARG). But MARG sensor can be easily affected by various disturbances, for example, vibration, external magnetic interference, and gyro drift. Optical flow sensor has the ability to extract motion information from image sequence, and thus, it is potential to augment three-dimensional attitude estimation for unmanned aerial vehicles. But the major problem is that the optical flow can be caused by both translational and rotational movements, which are difficult to be distinguished from each other. To solve the above problems, this article uses a gated recurrent unit neural network to implement data fusion for MARG and optical flow sensors, so as to enhance the accuracy of three-dimensional attitude estimation for unmanned aerial vehicles. The proposed algorithm can effectively make use of the attitude information contained in the optical flow measurements and can also achieve multi-sensor fusion for attitude estimation without explicit mathematical model. Compared with the commonly used extended Kalman filter algorithm for attitude estimation, the proposed algorithm shows higher accuracy in the flight test of quad-rotor unmanned aerial vehicles.

无人机的三维姿态估计通常基于磁力计，加速度计和陀螺仪（MARG）的组合。但是，MARG传感器很容易受到各种干扰的影响，例如，振动，外部磁干扰和陀螺仪漂移。光流量传感器具有从图像序列中提取运动信息的能力，因此，有可能为无人飞行器增加三维姿态估计。但是主要的问题是光流可能是由平移运动和旋转运动引起的，它们很难彼此区分开。为了解决上述问题，本文使用门控递归单元神经网络来实现MARG和光流量传感器的数据融合，从而提高无人飞行器三维姿态估计的准确性。所提出的算法可以有效地利用光流量测量中包含的姿态信息，并且可以在无需显式数学模型的情况下实现用于姿态估计的多传感器融合。与常用的扩展卡尔曼滤波算法进行姿态估计相比，该算法在四旋翼无人机飞行试验中具有更高的精度。