Fast discrimination by early judgment using linear classifier

Object detection involves classification of a huge number of detection windows obtained by raster scanning of the input image. For each detection window, a classifier trained with local features and a statistical learning method outputs a value for the target class. In this paper, we investigated the introduction of linear SVM approximate computation to object detection to increase the speed of raster scanning. We propose a method of fast discrimination by early judgment using linear classifier based approximation calculation. Doing so enables high-speed linear SVM classification by adaptively determining the number of bases required in the approximation calculations for the input detection window. Also, higher accuracy is attained in the object detection by representing the co-occurrence of binary-coded (B-HOG) forms of the HOG features that are used when doing the linear SVM approximating calculations. Evaluation experiments on human detection show that the proposed method is faster than using HOG features and linear SVM by a factor of 17 and improves the classification accuracy by about 6.1%.