Every Feature Counts: An Improved One-Stage Detector in Thermal Imagery
Yu Cao, Tong Zhou, Xinhua Zhu, Yan Ling Su · 2019
Detecting objects such as pedestrians and cars plays an indispensable role in autonomous driving. To be robust to the low-illumination situation such as nighttime, thermal cameras have become increasingly popular in autonomous driving system, and it is highly desirable to develop an accurate, robust algorithm concentrated on object detection in thermal imagery. In this work, we propose an DNN-based, one-stage detector namely ThermalDet. The main idea of ThermalDet is that since there doesn't exist lots of detailed visual characteristics (e.g. color and texture) in thermal images, features from low levels and high levels are equally important when performing detection task on them. The proposed detector inherits the architecture of RefineDet and further improves it. First, we design a dual-pass fusion block(DFB) to directly fuse features from all different levels. Then we add a channel-wise enhance module(CEM) to adaptively assign weights to different channels of the feature maps to achieve the best use of them. Experimental results on FLIR ADAS Dataset demonstrate that ThermalDet performs better than the state-of-art methods such as MMTOD-UNIT, MMTOD-CG.