LiDAR DNN based self-attitude estimation with learning landscape regularities

This paper presents an EKF (extended Kalman filter) based self-attitude estimation method with a LiDAR DNN (deep neural network) learning landscape regularities. The proposed DNN infers the gravity direction from LiDAR data. The point cloud obtained with the LiDAR is transformed to a depth image to be input to the network. It is pre-trained with large synthetic datasets. They are collected in a flight simulator because various gravity vectors can be easily obtained, although this study focuses not only on UAVs. Fine-tuning with datasets collected with real sensors is done after the pre-training. Data augmentation is processed during the training in order to provide higher general versatility. The proposed method integrates angular rates from a gyroscope and the DNN outputs in an EKF. Static validations are performed to show the DNN can infer the gravity direction. Dynamic validations are performed to show the DNN can be used in real-time estimation. Some conventional methods are implemented for comparison.