Motion LoD (Level of Detail) is a preprocessing technique that generates multiple details of captured motion by eliminating joints. This LoD technique is applied to movie, game or VR environments for the purpose of improving speed of crowd animation. So far, replacement techniques such as impostor and rigid body motion are widely used on real-time crowd, since they dramatically improve speed of animation. However, our experiment shows that the number of joints has a greater effect on the animation speed than anticipated. To exploit this, we propose a new motion LoD technique that not only improves the speed but also preserves the quality of motion. Our approach lies in between impostor and skeletal animation, offering seamless motion details at run time. Joint-elimination priority of each captured motion is derived from joint importance, which is generated by the proposed posture error equation. Considering hierarchical depth and rotational variation of joint, our error equation measures posture difference successfully and allows finding key posture of the entire motion. This motion analysis process contributes error reduction to the next motion simplification stage, where multiple details of motion are regenerated by the proposed motion optimization. In order to reduce the burden of optimization, all the terms of the objective function - distance, string, and angle error - are defined by joint-position vectors. In this aspect, a constrained optimization problem is formulated in a quadratic form. Thus, a sequential quadratic programming (SQP), a nonlinear optimization method, is suitable for resolving this problem. As the result of our experiment, the proposed motion LoD technique improves the animation speed and visual quality of simplified motion. Moreover, our approach reduces the preprocessing time and automates the whole process of LoD generation.