This thesis addresses the critical challenge of Ghost Object (GO) detection in Autonomous Vehicle (AV) perception systems. GOs, or False Positive (FP) object detections pose significant safety risks and hinder the reliable deployment of AVs. To tackle this issue, a labeled dataset of GOs was created through the manual annotation of real-world driving data obtained from a fleet of autonomous trucks operating in diverse conditions. Two distinct feature sets were derived from the trajectory data of annotated objects. The discriminative power of these feature sets was assessed and visualized using the Uniform Manifold Approximation and Projection (UMAP) Dimensionality Reduction (DR) technique, revealing inherent patterns and potential for distinguishing GOs from normal objects. A comprehensive evaluation of seven Machine Learning (ML) classification models and two Deep Learning (DL) approaches was conducted. The Adaboost model, trained on a specific feature set incorporating object lifetime, volume, velocity, acceleration, and distance metrics, emerged as the most effective ML model. While the DL approaches showed promise, their performance was constrained by the limited size of the labeled dataset and the models’ inherent complexity. A sensitivity analysis was performed to gain deeper insights into the underlying factors contributing to GO occurrences. This analysis revealed correlations between GO detections and specific environmental conditions, including high traffic density, daylight hours, and challenging road geometries such as tunnels and bridges.