Improving pedestrian detection

Pedestrian detection is an active problem in computer vision research, with applications in robotics, self-driving cars and surveillance. It involves generatting bounding boxes to indicate the location of every pedestrian in an input image. This paper proposes a method to augment a basic pedestrian detector with a Convolutional Neural Network. An implementation of the proposed algorithm was trained and tested on an ensemble of widely used pedestrian detection datasets, and achieved a 30% decrease in the FPPI (false positives per image) metric over an unaugmented HOG detector. However, it also increased the miss rate by 3.21%.