Efficient Attitude Estimation Through a Multilayer Perceptron Neural Network Based on Multi-Data Sensor Fusion Approaches

Abstract

Inertial sensor attitude estimation is essential for many applications, ranging from self-driving cars to tracking human movement. Ambient conditions, the presence of disturbances, and motion characteristics vary throughout application scenarios. Since modern attitude estimators do not generalize across different motion features and ambient conditions, their parameters must be changed for particular motion scenarios. In order to provide precise navigational solutions, global positioning systems (GPS) and IMUs are usually combined. In signal-challenged areas, the GPS receiver will not be able to provide accurate location information. The integration system will therefore become an IMU-independent system. In order to improve the performance of the attitude estimator, this paper describes a multi-data sensor fusion system that integrates IMU data using a multi-layer perceptron neural network. When estimating a vehicle’s attitude, one of the most important factors is the multi-combination of measurements from the MARG sensor. This problem cannot be solved by classical estimators; instead, a sensor fusion method needs to be created in order to produce reliable fusion outcomes. The authors suggest utilizing a multilayer perceptron model to execute a successful prediction in order to produce a highly exact attitude estimation. The tests conducted on the sequence demonstrate that the result produced surpasses the state-of-the-art, including Madgwick, Mahony, and complementary filters, when the right dataset is used for the orientation estimation.