Pedestrian Detection with Central-Line Heatmap Regression

The bounding box of a pedestrian is supposed to be in a rectangular shape with a normalized aspect ratio as it is used in annotation for pedestrian detection problem. However, detectors based on bounding-box regression suffer from hard negatives which are not accurate but contain parts of the human body. The main reason of this defect is that the detector is not highly aware of the central of a pedestrian. Motivated by this, we adopt a unique form of annotation which explicitly marks out the central of a pedestrian with a 1-pixel vertical line. We propose a novel one-stage detector which regresses a heatmap of lines using this annotation. The network of the detector constructs a hierarchy of feature maps and a built-in spatial module utilizes the spatial relations between points for enhancing the performance. Furthermore, a fast and effective CDF line-detector is presented for capturing salient responses in the heatmap to yield detections. Promising result is achieved on Caltech pedestrian dataset.