Multi-part-detector for human detection

Abstract

The paper proposes an capable approach of handling partial occlusion and local pose variation. Part detectors which contain position information for half of the sliding window are learned from the training data using the HOG feature and Adaboost. For each testing window, the response of each part detector is summed as a final response. With multi-part-detector approach which only need to compute gradient of the window once, better performance is achieved than whole window detector on the INRIA dataset.