Fast human detection via a cascade of neural network classifiers

In this paper, we build a cascade of neural network classifiers for fast human detection. The human object is represented by a collection of blocks. For each block, the histogram of orientated gradients feature is extracted and a neural network classifier is built as weak hypothesis. Then these hypotheses are selected sequentially by Gentle Adaboost and the cascade structure is used to speedup the detector. Compared to global linear SVM classifiers, the new method gets better performance on the INRIA human detection database at a much faster speed.