Fast Human Detection Using Deformable Part Model at the Selected Candidate Detection Positions

摘要

We integrate the classic deformable part models (DPM) with the object proposal approaches to achieve a fast and accurate human detection system. The proposed method avoids exhaustive sliding window search, which accelerating the detection speed and reducing the incorrect false positives. In this paper, EdgeBoxes and BING are selected as the candidate object proposal methods to generate the candidate detection positions for the DPM, because their good performance and fast speed. The DPM is only carried on the candidate locations selected by EdgeBoxes and BING for fast human detection. Experiments on PASCAL 2007 dataset for human detection show that the proposed method accelerates the detection speed and reduces the incorrect detections effectively, and EdgeBoxes is better than BING.