Autonomous drones require accurate navigation and localization algorithms to carry out their duties. Outdoors drones can utilize GPS for navigation and localization systems. However, GPS is often unreliable or not available at all indoors. Therefore, in this research, an autonomous indoor drone navigation model was created using a deep learning algorithm, to assist drone navigation automatically, especially in indoor corridor areas. In this research, only the Caddx Ratel 2 FPV camera mounted on the drone was used as an input for the deep learning model to navigate the drone forward without a collision with the wall in the corridor. This research produces two deep learning models, namely, a rotational model to overcome a drone's orientation deviations with a loss of 0.0010 and a mean squared error of 0.0009, and a translation model to overcome a drone's translation deviation with a loss of 0.0140 and a mean squared error of 0.011. The implementation of the two models on autonomous drones reaches an NCR value of 0.2. The conclusion from the results obtained in this research is that the difference in resolution and FOV value in the actual image captured by the FPV camera on the drone with the image used for training the deep learning model results in a discrepancy in the output value during the implementation of the deep learning model on autonomous drones and produces low NCR implementation values.
Read full abstract