Human detection based on curvelet transform and integrating heterogeneous features
Hong Wei Han, Youjian Fan · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2011
A method in Curvelet transformation and integrating heterogeneous features for human detection are proposed in this paper. The descriptor based on the second generation Curvelet transform (CTD) was proposed firstly, it concatenated the edge and texture feature vectors. To capture edge features, the statistic measures such as energy, entropy, standard deviation, max value and contrast computed from the blocks which is partitioned from the sub-bands of all the scales are concatenated. To get texture features, the lowest frequency sub-band coefficients were partitioned into overlapped blocks. Four co-occurrence matrixes were computed for each block. And some descriptors such as angular second-moment, contrast, correlation, sum of variance, sum of average and entropy are computed from the co-occurrence matrix, which are concatenated as the texture feature vector. And then the method integrating three feature extraction methods, such as Histogram of Oriented Gradient (HOG), Granularity-tunable Gradients Partition descriptors (GGP), and CTD, is proposed for human detection. Computational Cost Normalized classification Margin is used to determine the order of the feature to be evaluated. The experimental results on the basis of INRIA and MIT human database showed that CTD and integrating heterogeneous feature method increased the detection accuracy comparing to HOG and GGP.