摘要：

为解决布料与精细建模物体间碰撞检测速度慢、剔除率低等问题，提出简化模型的有向包围盒(OBB)算法和使用深度神经网络优化的连续碰撞检测（CCD）算法提高碰撞检测效率。在粗略检测阶段, 提出一种简化模型的OBB算法，对于精细建模物体使用二次误差度量的表面简化法对精细模型进行简化，将简化后的模型嵌入原模型中，并利用一种快速自适应的OBB算法对简化后的模型构建包围盒。对于布料模型，构建固定球形-轴向混合包围盒和碰撞检测包围盒树。在精确检测阶段，采用全连接深度神经网络学习滤波器剔除在粗略检测阶段未发生碰撞的碰撞对，使用训练后的最优深度神经网络模型优化连续碰撞检测算法。实验结果表明，使用简化程度为90%的模型构建的OBB可以完全替代原模型的OBB，所提简化模型的OBB算法与传统OBB包围盒算法及快速自适应包围盒算法相比，在耗时上分别缩短了约64.6%、35.8%。在布料与精细建模物体交互的场景下，使用深度神经网络优化的CCD算法比使用不同类型滤波器优化的CCD算法速度更快，耗时缩短了约7%~11%。

关键词: 碰撞检测, 布料模拟, 模型简化, 有向包围盒, 深度神经网络, 连续碰撞检测

Abstract:

To solve the limitations caused by slow speed and low culling rate of collision detection between cloth and finely modeled objects, an Oriented Bounding Box(OBB) algorithm with simplified model and a Continuous Collision Detection (CCD) algorithm optimized by deep neural network are proposed to improve the efficiency of collision detection.In the rough detection stage, an OBB algorithm of simplified model is proposed.For finely modeled objects, the surface simplification method of quadratic error measure is used to simplify the fine model.The simplified model is embedded into the original model, and a fast adaptive OBB algorithm is used to constructe bounding boxes for the simplified model.For the cloth model, the fixed spheric-axial hybrid bounding box and collision detection bounding box tree are constructed.In the precise detection stage, the fully connected deep neural network learning filter is used to eliminate the collision pairs that do not occur in the rough detection stage, and the trained optimal deep neural network model is used to optimize the continuous collision detection algorithm.The experimental results show that the OBB constructed by using the 90% simplified model can entirely replace the directed bounding box of the original model.Compared with the traditional OBB algorithm and the fast adaptive bounding box algorithm, the proposed OBB algorithm of simplified model algorithm requires 64.6% and 35.8% shorter time, respectively.In the case of interaction between cloth and finely modeled objects, the CCD algorithm optimized by deep neural network is faster than the CCD algorithm optimized by other different types of filters, and the time is shortened by approximately 7%-11%.

Key words: collision detection, cloth simulation, model simplification, Oriented Bounding Box(OBB), deep neural network, Continuous Collision Detection(CCD)