摘要
针对复杂场景下的交通目标分类识别难点,提出一种基于尺度不变特征转换(SIFT)与核稀疏表示的分类识别算法.该算法首先利用SIFT分别提取训练样本和待测目标局部特征信息,通过核方法将特征样本映射到核空间,构建过完备字典,最后通过待测目标在字典中的稀疏度与重构误差对交通目标类别进行判定.同时,分析了随机投影下的核稀疏表示分类与特征维数之间的关系.实验结果表明,与SVM、稀疏表示分类(SRC)相比,该方法增强了交通目标特征层的类判别能力,具有较好的识别率和鲁棒性.
A novel approach based on scale-invariant feature transform（ SIFT） and kernel sparse representation for traffic object recognition in complex traffic scenes is proposed in this paper. First,SIFT is introduced for feature extraction from samples and test targets,respectively. The features are mapping to the kernel space,then we construct an over-complete dictionary based on kernel sparse representation,traffic objects are recognized by computing sparsity and reconstruction residuals in the dictionary. We also analyze the relationship between recognition rate and dimensionality reduction of the SIFT descriptor using random projection. Experiment results show that the proposed approach enhances the class discriminant ability using traffic features with higher recognition preciseness and robustness in complex traffic scenes compared with SVM,SRC.
出处
《电子学报》
EI
CAS
CSCD
北大核心
2014年第11期2129-2134,共6页
Acta Electronica Sinica
基金
国家自然科学基金(No.61301027
No.61375015
No.11274226)
关键词
核稀疏表示
尺度不变特征转换
交通目标识别
压缩感知
随机投影
kernel sparse representation
scale-invariant feature transform（SIFT）
traffic object recognition
compressive sensing
random projection.
