<p>Although imitation learning from multiple human demonstrations has proven effective for robotic task learning, challenges remain in the setting of via-points and trajectory modulation in complex scenarios. This paper proposes a trajectory prediction and modulation method that integrates heuristic segmentation with a dual-layer kernelized movement primitive (KMP) framework. A reference trajectory is first generated using Gaussian Mixture Model-Gaussian Mixture Regression (GMM-GMR), after which the first layer KMP predicts the trajectory and performs heuristic segmentation. The second layer KMP then modulates trajectory segments near via-points, and the original predicted subtrajectories are replaced by the modulated ones to generate the final reproduced trajectory. In addition, adaptive parameter optimizers are designed for both GMM-GMR and the first layer KMP to automatically determine the number of Gaussian components and kernel parameters, reducing reliance on empirical parameter tuning. Experimental results show that the proposed method generates smoother trajectories with higher prediction accuracy than mainstream movement primitive methods. In obstacle avoidance tasks, the proposed method preserves more movement primitive features while generating smoother and more natural avoidance trajectories. Overall, the method achieves a good balance between trajectory accuracy, smoothness, and computational efficiency, providing a robust solution for robotic motion planning in dynamic environments.</p>