Double Phase Pedestrian Detection with Minimal Number of False Positives per Image
Masoud Afrakhteh, Miryong Park · 2017
In this paper, we investigate pedestrian detection from visible color imagery in a manner that reduces the number of false positives (FPs) per frame. We propose a simple way to eliminate many of such unwanted FPs that usually result from applying a state-of-the-art pedestrian object detector to a visible color image. The symmetric structure of pedestrian bodies can often be a good clue to distinguish them from non-pedestrians. Hence, each and every individual detected object (bounded box) in the initial step of detection is used to reconstruct some new symmetric-looking objects. If any of these objects are detected, they are confirmed to be a person; otherwise, they are omitted from the very first list of detected objects. Experimental results show how these FPs are removed using such a simple technique while maintaining the previous low miss rate.