基于高阶类别相关先验知识的室外场景语义分割研究

发布时间：2018-04-19 20:03

本文选题：语义分割 + 点云　；参考：《南京大学》2017年硕士论文

【摘要】：三维室外场景的语义分割算法涉及到分割和分类两个方面,它的目标是将三维点云数据分割成物体,并将物体正确地分类。有效的语义分割能够帮助计算机理解三维室外场景的内容,同时为虚拟现实,无人驾驶和城市建模等应用提供巨大的帮助。目前,三维语义分割算法的一个方向是先分割再分类,但是这样容易在多个步骤中产生累积误差;另一个方向是利用条件随机场同时进行分割与分类,并且能够利用一些先验知识来优化结果。但是这些方法中的条件随机场要么只依赖于二阶项,要么它的高阶项所依赖的先验知识过于单一和简单,无法刻画复杂室外场景中的先验知识。本文提出了一种基于类别相关先验知识的高阶条件随机场模型。该模型依赖于高阶条件随机场,更关键的是能够针对不同类别物体的先验知识,产生对应的高阶能量函数。我们的方法有效地描述三维室外场景中复杂的统计信息,以及不同类别物体先验知识之间的区别。我们将该模型至于能量函数最小化的框架之中,并且提出具体的一阶,二阶和基于类别相关先验信息的高阶能量函数来进行三维室外场景的语义分割。提出了具体的能量函数之后,我们探讨了能量函数的优化问题。对于二阶能量函数,它的优化算法比较成熟和高效,例如graph Cuts,和loopy belief propagation。但是对于高阶能量函数,它的优化问题则是非常具有挑战性的。本文采用了基于线性规划和最大流最小割的原始对偶能量优化算法,有效地解决了高阶能量函数的优化问题,最终获得我们三维室外场景的语义分割结果。为了验证我们算法的有效性,我们在一个非常有挑战性的雷达数据集上测试我们的算法,并提供定量和定性的结果。相比于其他方法,我们实验结果的平均F1得分达到0.82,高于其他最新方法的0.73的F1得分。
[Abstract]:The semantic segmentation algorithm of 3D outdoor scene involves two aspects: segmentation and classification. Its goal is to divide 3D point cloud data into objects and classify objects correctly. Effective semantic segmentation can help the computer to understand the content of 3D outdoor scene and provide great help for applications such as virtual reality, driverless and urban modeling. At present, one direction of 3D semantic segmentation algorithm is segmentation and classification, but it is easy to generate cumulative errors in multiple steps, the other is to use conditional random fields to segment and classify simultaneously. And can use some prior knowledge to optimize the results. However, the conditional random fields in these methods are either dependent on the second order terms only, or their higher order terms depend on a priori knowledge which is too simple and simple to depict the prior knowledge in complex outdoor scenes. In this paper, a higher order conditional random field model based on class dependent prior knowledge is proposed. The model depends on higher order conditional random fields, and more importantly, it can generate the corresponding higher order energy function for the prior knowledge of different kinds of objects. Our method effectively describes the complex statistical information in 3D outdoor scenes and the difference between prior knowledge of different classes of objects. We take the model as the framework of minimizing the energy function, and propose specific first-order, second-order and higher-order energy functions based on class dependent prior information to segment 3D outdoor scenes. After putting forward the specific energy function, we discuss the optimization problem of the energy function. For second-order energy functions, its optimization algorithms are more mature and efficient, such as graph cuts and loopy belief propagation. But for higher order energy function, its optimization problem is very challenging. In this paper, the original dual energy optimization algorithm based on linear programming and maximum flow minimum cut is used to solve the optimization problem of high order energy function effectively. Finally, the semantic segmentation results of our 3D outdoor scene are obtained. To verify the effectiveness of our algorithm, we test our algorithm on a very challenging radar dataset and provide quantitative and qualitative results. Compared with other methods, the average F1 score of our experiment is 0.82which is higher than the 0.73 F1 score of other latest methods.
【学位授予单位】：南京大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP391.9

【相似文献】