Pedestrian Avoidance in Construction Sites

This paper proposes a pedestrian detection pipeline consisting of an Intel RealSense R200 camera used for input to a Single Shot Multibox Detector based neural network for pedestrian detection. Unlike prior research the proposed system uses a neural network for detection, then establishes the distance of a pedestrian from the camera using stereoscopy in realtime using commonly available low cost hardware. Two networks are presented for pedestrian detection, the Single Shot Multibox Detector (SSD) 512, and the InceptionV4 network which has been modified to use the SSD design for pedestrian detection. Both networks are trained on a dataset created from footage captured at a Fulton Hogan site. The SSD InceptionV4 network achieves 1% mAP on the PASCAL VOC 2007 test dataset and 51% mAP on the Fulton Hogan dataset. The SSD 512 network achieves 93% mAP on the Fulton Hogan dataset and 66% mAP on the PASCAL VOC 2007 test dataset.