Sparsity-motivated multi-scale histograms of oriented gradients feature for SRC

In order to recognize targets accurately from the low-quality images obtained from unmanned system, sparse representation based classification (SRC) method using sparsity-motivated gradient feature was proposed. The multi-scale histograms of oriented gradients (HOG) feature was used as an original feature, whose dimension was reduced by a non-adaptive random projection method. A very sparse measurement matrix was adopted to preserve the structure of multi-scale HOG feature space efficiently. The sparse representation was obtained via i
₁
-norm minimization, and the least reconstruction error was used as recognition principle. Experiment results again Comanche FLIR data set show that, the proposed method can raise the recognition rate by 2% compared with the state of art methods.