Dilated Light-Head R-CNN using tri-center loss for driving behavior recognition

作者：

Highlights：

• A novel approach is proposed to recognize driving behavior by detecting action specific parts.

• Dilated convolution is used to extract feature of small objects, balancing resolution and receptive field.

• A position-sensitive RoI alignment method is proposed to improve the perception of small objects.

• The tri-center loss enforces similarity between intra-class features and difference between inter-class features.

摘要

•A novel approach is proposed to recognize driving behavior by detecting action specific parts.•Dilated convolution is used to extract feature of small objects, balancing resolution and receptive field.•A position-sensitive RoI alignment method is proposed to improve the perception of small objects.•The tri-center loss enforces similarity between intra-class features and difference between inter-class features.

论文关键词：Driving behavior,Dilated convolution,Tri-center loss,Positive-sensitive RoI alignment

论文评审过程：Received 27 April 2019, Revised 12 August 2019, Accepted 14 August 2019, Available online 23 August 2019, Version of Record 17 October 2019.

论文官网地址：https://doi.org/10.1016/j.imavis.2019.08.004