FPPNet: A Fixed-Perspective-Perception Module for Small Object Detection Based on Background Difference
Wentao Liu, Bin Zhou, Zhangyu Wang, Guizhen Yu, Songyue Yang
Abstract
A roadside sensing unit can provide over-the-horizon perception information for autonomous vehicles due to its high perception perspective. However, numerous challenges need to be overcome such as the missing detection of small objects and occluded objects. To this end, this study proposed a fixed perspective perception (FPP) module, which considered background subtraction and a fixed camera for small object detection. The proposed FPP module was divided into two parts: a grayscale background subtraction (GBS) submodule and a background–current image fusion (BCF) submodule. Specifically, the GBS submodule introduces background spatial information into a current frame, and the BCF submodule combines feature maps of a current frame and background by using channel attention. Moreover, we designed an object detection network called FPPNet which uses the FPP module to facilitate small object detection. Experimental results demonstrate that the FPPNet achieved 39.8% average precision small ( <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${\text {AP}}_{{S}}$ </tex-math></inline-formula> ) and 65.7% AP in a Dair-V2X-I dataset. Futhermore, we conducted an extension of the FPP module to mainstream object detection networks such as CenterNet, Faster-Rcnn, and RetinaNet. Experimental results show that the proposed module can effectively improve small object detection accuracy of the networks mentioned earlier.