Real Time Pedestrian Detection-based Faster HOG/DPM and Deep Learning Approaches

Abstract

The work presented aims to show the feasibility of scientific and technological concepts in embedded vision dedicated to the extraction of image characteristics allowing the detection and the recognition/localization of objects. Object and pedestrian detection are carried out by two methods: 1. Classical image processing approach, which are improved with Histogram Oriented Gradient (HOG) and Deformable Part Model (DPM) based detection and pattern recognition. We present how we have improved the HOG/DPM approach to make pedestrian detection as a real time task by reducing calculation time. The developed approach allows us not only a pedestrian detection but also calculates the distance between pedestrians and vehicle. 2. Pedestrian detection based Artificial Intelligence (AI) approaches such as Deep Learning (DL). This work has first been validated on a closed circuit and subsequently under real traffic conditions through mobile platforms (mobile robot, drone and vehicles). Several tests have been carried out in the city centre of Rouen in order to validate the platform developed.