感知与规避技术中的入侵目标检测的特征选择

Image Feature Selection Mechanism for Intruder Detection in Vision-based Sense and Avoid System

论文针对基于视觉的感知与规避技术中的入侵目标检测,提出了一套稀疏表示框架下的图像特征选择机制。基于稀疏编码和空间金字塔匹配算法(sc-SPM)的低层特征描述子常用的是方向梯度直方图(HOG)特征和尺度不变特征转换(SIFT)特征,而论文通过对在复杂背景下不同天气情况的入侵目标检测结果的查全率(recall)曲线来比较这两种特征描述子性能,最后选择性能最好的特征描述子作为sc-SPM特征提取算法的底层特征。实验结果表明,SIFT特征描述子更能适用于多种不同天气情况并且具有更好的鲁棒性。

In this paper,an image feature selection mechanism under the framework of sparse representation is proposed forintruder detection in vision based on sense and avoid system. Based on the framework of Spatial Pyramid Matching using Sparse Cod-ing,the low-level feature usually adopts the Histogram of Oriented Gradient(HOG)feature descriptor or the Scale-Invariant Fea-ture Transform(SIFT)feature descriptor. In this paper,recall of the intruder detection results is used to compare the HOG featuredescriptor and the SIFT feature descriptor under different weathers with complex background. The best-performing feature descrip-tor is chosen as the low-level feature of sc-SPM method. The experimental results show that the SIFT feature descriptor is more suit-able for a variety of different weathers and has better robustness.