Investigation of Adaptive Kalman Filter Techniques for Sensor Fusion Applications

Abstract

Sensor fusion is a concept encountered in many different disciplines today. In its simplest definition, this concept aims to combine information from different sources using appropriate methods. In this thesis, a simulation environment and experimental setup utilizing LiDAR and a camera have been established. A closed-loop model was employed in the experimental setups to test multi-sensor algorithms. Initially, a superposition-based tracking behavior was examined using the closed-loop model. Using this method, sensor data was fused by weighting it with a constant value. Subsequently, a Kalman filter was utilized to enhance state estimation under different conditions. The error covariance provided by the Kalman filter was used for weighting. Independent of the general noise conditions of the Kalman filter, the system's performance was examined using different measurement noise estimation algorithms found in the literature. Thus, the error resulting from the position estimation of the sensors could be dynamically determined. By dynamically adjusting the weights, an effective result was achieved in cases where environmental noise varied, and as a consequence, the system's tracking performance was improved.