| Abstract (Eng): | The thesis proposes a method for ship motion prediction which is optimized to guide the landing of autonomous drones on unstable surface vessels. With this prediction, the drone can anticipate movements of the landing platform to reduce the landing impact and ensure the safety of both the drone and its surroundings. The target vessel is equipped with an onboard inertial measurement unit (IMU) motion sensor and a front-facing camera to facilitate the predictions. Based on the six degrees of object motion freedom, pitch and roll were chosen as primary prediction targets as they are most influential towards the stability of the landing platform. In future work, this could be expanded to include other parameters such as heave. Five deep learning neural network architectures were designed as prediction models in order to find an optimal architecture for the problem. Four out of the five models were capable of predicting multi-value sequential outputs with varying accuracy. It was found that the use of Long-Short Term Memory (LSTM) networks improved prediction accuracy. Additionally, the use of wave images resulted in the overall highest accuracy. However, performance decreases when the images contain obscurities. Models processing images as input also suffered from high latency. Based on empirical results, an encoder-decoder LSTM network architecture is proposed as the most optimal architecture. This model is capable of making one-minute-long predictions for pitch and roll with consistent accuracy. Tests resulted in an average prediction error of around two degrees and a latency of 155ms. The model uses previous pitch and roll measurements to make predictions which also omits its dependency on image clarity. All tests were performed in a simulated environment. |
|---|