Object bounding box-critic networks for occlusion-robust object detection in road scene

Abstract

Object detection in a road scene has received a significant interest in the research field of developing autonomous vehicles and automatic road monitoring systems. However, object occlusion problems frequently happen in generic road scene. Due to the occlusion problem, previous object detection methods have limitations that they could not detect objects correctly. In this thesis, we propose a novel object detection network aiming to an occlusion robust method. To effectively detect object even in occlusion cases, the proposed network mainly consists of two parts

1) Object bounding box (OBB)-critic network which handle occlusion early at feature map encoded from input image. 2) RoI object bounding box-critic network which handle occlusion at the RoI feature predicted in the Region Proposal Network (RPN). Two OBB-critic networks are trained by an adversarial learning. Comprehensive experimental results on a KITTI dataset showed that the proposed object detection network outperformed state-of-the-art object detection methods.