基于尺度不变特征和位置先验的行人检测算法
Pedestrian Detection Algorithm Based on Scale Invariant Features and Prior Position Information

基于相邻和非相邻特征(NNNF)行人检测算法, 提出了一种方法来解决行人特征对尺度变化敏感的问题以及窗口误检的问题．首先, 在NNNF基础上, 设计了一种尺度不变的特征．由韦伯定理启发, 该特征表示为两个相邻或非相邻区域的差分值与这两个区域特征和的比值．这种新的特征具有很强的尺度不变性．此外, 还提出了基于行人位置先验的上下文信息, 作为一种简单有效的后处理方法．在行人场景中, 行人的高度与位置存在一定的映射关系．利用SVM(support vector machine)训练了行人高度关于行人位置的回归模型．该模型能有效地滤除那些行人高度与位置信息不符合回归模型的检测窗口．实验表明, 相比于NNNF-L2和NNNF-L4, 本文提出的方法在Caltech数据库的检测性能分别提高了2.90% 和2.28% ．同时, 本文提出的方法也在所有基于非深度网络的行人检测方法中具有最好的检测性能, 平均漏检率为14.56% ．
Based on neighbouring and non-neighbouring features(NNNF)，a method was proposed to address challenges in pedestrian detection，such as features being sensitive to scale changes and the existence of a large number of false positives. First，based on NNNF，a scale invariant feature was proposed. This feature is computed as the ratio of difference to sum between two neighbouring or non-neighbouring regions’ local feature value，inspired by the Weber law in experimental psychology. The new feature is strong scale invariant. Second，a contextual information based on the prior position information was proposed，and this method serves as simple and effective post-processing approach. In the scene of pedestrians，there is some function relationship between the height and position of pedestrian. The support vector machine(SVM)was utilized to train a regression model of height in relation to position. And this regression model can effectively filter out those false positives whose height and position are inconsistent with the regression model. Furthermore，experimental results on Caltech pedestrian dataset show that the proposed method gives an average 2.90% and 2.28% improvement over the NNNF-L2 and NNNF-L4，respectively. Most importantly，among all the methods without using CNN，the proposed method achieves state-of-the-art performance(i. e.，14.56% average miss rate on Caltech dataset)