非重叠低维度梯度方向直方图

Low-Dimensional Histogram of Oriented Gradients with Non-Overlapping Scheme

作为衍生于尺度不变特征变换的特征描述,梯度方向直方图(HOG)在人体检测、手势识别、人脸识别、场景分类等方面得到广泛应用.但HOG的特征维数高,导致维数灾难和大计算量.文中发现HOG特征的高维度源自它需在众多重叠块中计算直方图.虽然重叠块机制对特征的鲁棒性有积极作用,但也导致信息冗余.为去除冗余信息并降低特征维数,从直方图归一化入手,提出非重叠式梯度方向直方图.所提方法的维数降低为传统方法的1/3.在人手和人体检测上的实验表明,该方法不仅物体检测速度得到显著提高,检测准确度也得到改善.

As a derivation version of scale-invariant feature transform （SIFT）, histogram of oriented gradients （HOG） is widely used in human detection, gesture recognition, face recognition, scene classification, etc. However, the high dimension of the HOG feature vector leads to the curse of the dimensionality and high computation complexity. In this paper, it is found that the high dimension of HOG feature vector results from computing histograms of overlapping blocks. Though overlapping block is useful for enhancing the robustness, it leads to redundant information. To reduce the redundant information and the number of features as well, a non-overlapping version of HOG is proposed. The dimensions of the proposed method are 1/3 of those of traditional ones. The experimental results on palm and human detection demonstrate the efficiency and effectiveness of the proposed method.