Perpustakaan Digital ITB

GPS signals cannot penetrate buildings in indoor environments, making it difficult for drones to locate their position. Using external visual sensors with object detection models offers an alternative approach. However, these sensors have drawbacks, including the risk of false detections due to prioritizing computation speed over detection accuracy. Moreover, existing research focuses more on positional determination rather than developing autonomous guidance capabilities utilizing external visual sensor data. To address these challenges, this study aims to determine the drone’s position in real-time using a fast-computing computer vision algorithm. A Kalman Filter algorithm is then used to enhance the resilience of the external visual sensors to a false detections. The study integrates the Kalman Filter’s positional estimation data into a waypoint-following algorithm for autonomous indoor navigation. The research involves a data collection of camera parameters and training datasets for the computer vision model, flight testing, analysis, and conclusion The external visual sensor uses curve fitting and camera model equations to convert pixel coordinates into local coordinates, achieving top performance with errors of 3.48% and 2.29% for X and Y axis respectively. Position determination is further enhanced using a sensor fusion algorithm, combining data from external visual sensors, accelerometers, and optical flow sensors via an Adaptive Kalman Filter. The Adaptive Kalman Filter dynamically adjusts measurement and process noise covariances using a forgetting factor and compares the estimation data with previous estimations to increase robustness during prolonged false detections. Subsequent flight tests demonstrate the integrated system’s ability to navigate to waypoints while effectively reducing the impact of false detections, but the position prediction during false detections depends on the quality of its onboard sensors.