红外图像中基于似物性与稀疏编码的行人检测

Pedestrian Detection Based on Objectness and Sparse Coding in a Single Infrared Image

行人检测是计算机视觉的经典问题。针对红外图像中的行人检测问题,提出了一种基于似物性和稀疏编码及空间金字塔特征提取的行人检测方法。首先,针对红外图像的特点,利用基于频域残差的显著性分析方法得到红外图像的显著图,在此基础上提出了一种似物性计算方法,进而得到不同区域的似物度得分,并根据得分提取出感兴趣区域;其次,以尺度不变特征转换为基础,将稀疏编码和空间金字塔算法应用于非监督特征学习实现对感兴趣区域的特征提取;最后,利用线性支持向量机构建分类器实现对图像中每个感兴趣区域的行人检测。实验结果验证了本文提出的感兴趣区域提取算法和针对单幅红外图像行人检测算法的有效性。

Pedestrian detection is a classic issue of computer vision. For the pedestrian detection problems in a single infrared image, this paper proposes a pedestrian detection method based on objectness, sparse coding and spatial pyramid matching. The algorithm can be divided into three phases. Firstly, the saliency map is computed based on spectral residual, and the paper presents an objectness score computation based on saliency map and selects regions of interest according to the score of different sub-windows. Secondly, scale-invariant feature transform, sparse coding and spatial pyramid matching are used to extract the feature vectors of the regions of interest. Finally, linear support vector machine is used to build a classifier and detect pedestrian in each region of interest. The experimental results verify the effectiveness of objectness score computation and the proposed algorithm for infrared images.