Vision based People Detection in Power Substations

Abstract

In reality, it is desired to reliably detect unauthorized entries for prohibited areas such as teleoperated high voltage power substations. To this end, this paper presents a vision based people detection method in power substations. More than 42,000 images were collected from the COCO dataset, Youtube and a real 220kV power substation for training models. Deep learning models EfficientDet-D1 and YOLOv5-m were employed for transfer learning. Experimental results show that the EfficientDet-D1 and YOLOv5-m can recognize people in 220kV power substations with mAP of 63.2% and 88.6%, respectively.