Hand-Crafted Features vs Deep Learning for Pedestrian Detection in Moving Camera

Detecting pedestrians and other objects in images taken from moving platforms is an essential task needed for many applications such as smart surveillance systems and intelligent transportation systems. However, most detectors in this domain still rely on handcrafted features to separate the foreground objects from the background. While these types of methods have presented good results with good response times, they still have some weaknesses to overcome. In recent years, alternative object detection methods are being proposed, with deep learning based approaches rising in popularity thanks to their promising results. In this paper, we propose two pedestrian detectors for use in images taken from a moving vehicle: The first detector uses a block matching algorithm and handcraft features for pedestrian detection, and the second uses a Faster R-CNN deep detector. We also compare both systems' performances to other state-of-the-art pedestrian detectors. Our results show that although handcraft feature-based approach achieves good results within acceptable detection times, it suffers from a high false positive rate. However, we found that Faster R-CNN detector performs better in terms of precision and recall, but these improved results come at a cost of detection time.