earticle

영어

Sliding window (SW) technique is one of common paradigms employed for object detection. However, the computational cost of this approach is so expensive because the detection window is scanned at all possible positions and scales. To overcome this problem, we propose a compact feature together with fast recursive coarse-to-fine object localization strategy. To build a compact feature, we project the Histograms of Oriented Gradient (HOG) features to linear subspace by Principal Component Analysis (PCA). We call this feature as PCA-HOG feature. The exploitation of the PCA-HOG feature not only helps the classifiers run faster but also still maintains the accuracy. In order to further speeding up the localization, we propose a recursive coarse-to-fine refinement to scan image. We scan image in both scale space and multi-resolution space from coarsest to finest resolutions. Only the best obtained hypothesis from the coarser resolution could be passed to finer resolution. Each resolution has its own linear Support Vector Machine (SVM) classifier and PCA-HOG features. Evaluation with INRIA dataset shows that our method achieves a significant speed-up compared to standard sliding window and original HOG feature, while even get higher detection accuracy.