Dual Guidance Of Optical Flow And Decoupled Attention For Infrared Video Object Detection Network.

Infrared video object detection is a task to detect moving objects in infrared video. The traditional object detection algorithm often ignores the motion information of the target, and often miss detection and false detection of small objects and objects in complex backgrounds. This paper proposes a Flow Attention Dual-Guided Yolo (FADG-YOLO) algorithm. It is used to solve the problems such as difficult detection of moving targets, small and weak targets, and complex backgrounds. Specifically, our algorithm includes two modules, FlowFocus (FF) and Decoupled Attention (DA). FlowFocus is a module that calculates optical flow, aligns motion features with original image features and fuses them. It is used to introduce motion information into the model. Decoupled Attention is a module that enhances the feature responses of small objects and difficult-to-recognize objects in complex backgrounds. Based on the two modules of FlowFocus and Decoupled Attention, this paper constructed the FADG-YOLO algorithm for infrared video object detection. The result from extensive experiments on the ICCV2021 Anti-UAV Challenge Dataset show that our proposed algorithm has better performance than many popular object detection models.