Accurate and flexible trajectory generation and optimization are essential for autonomous driving decision and planning system to handle complex traffic and diverse driving styles. In this paper, the decision algorithm for lane-change scenarios in Urban Navigation Pilot system of a mass-production vehicle is enhanced. 1) A reference trajectory generator based on grid map construction and spatio-temporal sampling is developed. The Iterative Linear Quadratic Regulator (iLQR) is then employed, along with engineering improvements, to optimize the selected reference trajectory from the generator. For collision preclusion and adherence to traffic laws, constraints are incorporated into iLQR through augmented Lagrange approach. Overall, the semantic decision, along with the corresponding reference trajectory and boundaries, can be jointly generated and output by the decision module. 2) Compared to other algorithms, iLQR demonstrates superior performance in lane-change execution and achieves lower time cost. Experiment and simulation results demonstrate that the iLQR-based decision module exhibits comprehensive improvements, compared to the baseline approach. Notably, refinement of the spatio-temporal corridor leads to a reduction in the proportion of deceleration-free lane-changes from 55.9% to 26.5%. In emergency scenarios, iLQR algorithm outperforms the original in 60% of cases. Under extreme conditions—such as low-speed driving behind a stationary vehicle—the minimum required distance for a successful lane-change is reduced from 8.5m to 6.5m, indicating enhanced maneuverability. 3) A GPU-based parallel implementation of iLQR algorithm is developed. Significant improvement is observed when the forward pass linear search is parallelized, reducing the time cost by 38.1%.