Predicting pedestrian motion in indoor enviroment is vital for robot navigation and human-robot interaction. However, it is challenging for traditional filter-based method due to the inherent unpredictability of human movement. This paper introduces a novel Lidar-Vision based framework that enhances prediction accuracy by incorporating predictive motion priors. The proposed method dynamically adapt the filter’s process noise by analyzing real-time human pose features and environmental map constraints. This prior-aware adaptation enables superior tracking of non-linear maneuvers like turns and avoidance behavior. Experiments demonstrate substantial improvements in prediction accuracy and robustness compared to baseline methods on challenging indoor trajectories.