Research on the Detection Technique of Situation Elements in Obscure Overlapping Scenes
Jinlong Liu, Kangda Cheng · Wireless Communications and Mobile Computing · 2022
In recent years, some scholars have proposed to apply single‐stage target detection algorithms such as YOLO (You Only Look Once) to situational element detection, but the traditional YOLO algorithm is to treat the target detection process as a regression problem, which cannot distinguish well between overlapping objects and has defects such as less accurate bounding boxes and hard to distinguish objects from the background, and it is difficult to cope with problems such as the higher overlap of targets to be detected and stronger target camouflage ability in obscured overlapping scenes. In this paper, we propose to add the attention module CBAM to the backbone network of the YOLOv3 model, to construct a SEDNet with high accuracy and good robustness for situational element detection, and to apply it to the situational element detection in occlusion overlapping scenes. We use SEDNet to classify and localize ten elemental targets, respectively. The analysis of experimental results shows that the SEDNet target detection model can complete element detection in complex environments with strong target camouflage, achieve end‐to‐end detection, and lay the technical foundation for the formation of complete situational awareness.