Pedestrian detection based on multifractal analysis

This paper proposes a new multifractal analysis based feature representation for object representation. Multiple Fractal Dimensions (MFD) are calculated to describe the distribution of fractal dimensions measured on a finite number of point sets extracted from the image. The proposed MFD feature is theoretically proven to be invariant to articulations, which is a desirable characteristic for faces and pedestrian due to the existence of expressions, posture and illumination variations. The new object representation is extensively evaluated on pedestrian detection problem. The experiments INRIA pedestrian databases show that our method achieves a much better performance than baseline methods in terms of recognition rates.