Investigate feasibility of utilising a neural-networked set of inertial measurement units to compensate for variations in motion of a commercial radio-controlled vehicle in a dryland agricultural context : A thesis submitted in partial fulfilment of the requirements for the Degree of Master of Applied Science at Lincoln University

Abstract

Image stabilisation is desired for efficient identification of objects in the path of a self-driving vehicle. The gyroscope and accelerometer of an inertial measurement unit (IMU) can be used to derive the movement of a vehicle, which can then be used by a rotation matrix to compensate for this movement, but a gyroscope has inherent “drift” errors, and while the accelerometer of an IMU is more accurate, it has a slower response time, which reduces the detection rate. Various methods have been proposed to compensate for these sensor limitations. Kalman filters are often used in industry to fuse gyroscope and accelerometer data to reduce the effects of drift, noise, and other gaussian-based errors but these are computationally intensive for the sort of lightweight processor that a radio-controlled car could be expected to power. A complementary filter such as Madgwick’s is a simpler and less processor-intensive solution with claims that the method is just as accurate. Both of these approaches are applied on a single IMU. Averaging multiple IMUs has been investigated and offers slight improvements. Well-trained neural networks also offer IMU compensation but are computationally and time expensive to train to generate a model however the application of previously-trained models is less intensive and is becoming common as processor power improves. There have been no investigations of using a neural network on multiple IMUs as of this time. This project investigates if using a neural network of multiple IMUs reduces errors and enhances performance compared to a single IMU. The Kalman filter is used as baseline control data and three neural network models (MLP, NARXNET and RBF) are compared against each other and a Madgwick complementary filter to investigate if using a neural network of multiple IMUs reduces errors and enhances performance compared to a single IMU in the context of establishing Euler angles of roll and pitch movement to stabilise a video feed of a consumer-level camera on a moving commercial off-the-shelf radio control vehicle. It is demonstrated that there is no statistically significant advantage in using multiple IMUs if these are kept in the same horizontal plane, that a minimum of a three-layer MATLAB NARXNET filter provides the equivalent accuracy of a Kalman filter with similar processing times, and that the Madgwick IECF6 complimentary filter, the radial basis factor neural network and multi-level perceptron neural network models are not fit for this purpose.