Semantic Fusion-based Pedestrian Detection for Supporting Autonomous Vehicles

EasyChair Preprint 3783

Abstract

To increase traffic safety and transportation efficiency, adopting intelligent transportation systems (ITS) has become a trend. As an important component of ITS, one essential task of autonomous vehicles is to detect pedestrians accurately, which is of great significance for improving traffic safety and building a smart city. In this paper, we propose an anchor-free pedestrian detection model named Bi-Center Network (BCNet) by fusing the full body center and visible part center for each pedestrian. Experimental results show that the performance of pedestrian detection can be improved with a strengthened heatmap, which combines the full body with the visible part semantic. We compare our BCNet with state-of-the-art models on the CityPersons dataset and the ETH dataset, which shows that our approach is effective and achieves a promising performance.

Keyphrases: Convolution Neural Network, Intelligent Transportation System, autonomous vehicle, pedestrian detection

@booklet{EasyChair:3783,
  author    = {Mingzhi Sha and Azzedine Boukerche},
  title     = {Semantic Fusion-based Pedestrian Detection for Supporting Autonomous Vehicles},
  howpublished = {EasyChair Preprint 3783},
  year      = {EasyChair, 2020}}