Real Time Pedestrian Detection Using Modified YOLO V2

Object detection is an important task in autonomous driving. Identifying an object and localizing it in the given 2D space is comparatively easier task for humans but for an autonomous vehicle, the case is not the same. Recently, many algorithms were proposed to mitigate these issues out of which YOLO (You Only Look Once) is out of the box in a sense that it can run faster without compromising the accuracy. Pedestrian detection is the most difficult and critical application in automotive perspective since any error can be fatal. In this paper, pedestrian detection and localization is mainly focused using our modified YOLOv2 architecture (Model H). Our architecture (Model H) results were compared with the YOLOv2 baseline model's results for well-known INRIA person dataset and our approach yield better results. Also implemented it in embedded system hardware using NVIDIA Jetson TX1 and Zed Stereo camera with 30 fps detection rate. Apart from detection, depth is also estimated using the ZED camera and depth map is created.