基于多条件随机场模型的图像3D空间布局理解

图像3D空间布局理解在自动驾驶系统以及目标识别中扮演着重要的角色.本文提出一种基于多条件随机场模型集成的图像3D空间布局理解算法.首先，基于多次图像分割生成多个不同尺度的超像素图像；然后，结合LBP表面纹理特征、LM滤波器组获得的方向纹理特征、颜色特征以及图像中超像素的位置和形状特征，建立各尺度的超像素图像中超像素的特征表达；最后，为各尺度的超像素图像分别构建相应的条件随机场模型，并应用D-S证据合成理论对多个条件随机场模型的推断结果进行集成，实现对图像3D空间布局的理解.在公共数据集GC和KITTI Layout上的实验结果表明，同已有算法相比，本文提出的算法提高了图像3D空间布局理解的准确率.

3D Spatial Layout Understanding from Image Based on Multiple CRFs

3D spatial layout understanding from images plays an important role in the autonomous driving and object recognition.This paper proposes a 3D spatial layout understanding algorithm based on multiple CRFs.Firstly,multiple different scales super-pixel image are generated based on the multiple segmentation.Then,the feature of super-pixel are constructed based on LBP surface texture feature,orientation texture feature from LM filters,color feature,and location and shape feature of super-pixels in the image.Finally,the CRF model on every scale super-pixel image is built,the Dempster-Shafer theory of evidence is used to integrate the inference result of multiple CRF models and the 3D spatial layout understanding from an image is realized.The experiments on the public database Geometric Context and KITTI Layout demonstrate that the algorithm proposed in this paper improves the average accuracy of 3D spatial layout understanding comparing to the existing state-of-art.