BGI-YOLO: Background Image-Assisted Object Detection for Stationary Cameras

This paper proposes a method enhancing the accuracy of object detectors by utilizing background images for stationary camera systems. Object detection with stationary cameras is highly valuable across various applications, such as traffic control, crime prevention, and abnormal behavior detection. Deep learning-based object detectors, which are mainly used in such cases, are developed for general purposes and do not take advantage of stationary cameras at all. Previously, cascade-based object detection methods utilizing background have been studied for stationary camera systems. These methods typically consist of two stages: background subtraction followed by object classification. However, their object detection performance is highly dependent on the accuracy of the background subtraction results, and numerous parameters must be adjusted during background subtraction to adapt to varying conditions. This paper proposes an end-to-end object detection method named BGI-YOLO, which uses a background image simply by combining it with an input image before feeding it into the object detection network. In our experiments, the following five methods are compared: three candidate methods of combining input and background images, baseline YOLOv7, and a traditional cascade method. BGI-YOLO, which combines input and background images at image level, showed a detection performance (mAP) improvement compared to baseline YOLOv7, with an increase of 5.6%p on the WITHROBOT S1 dataset and 2.5%p on the LLVIP dataset. In terms of computational cost (GFLOPs), the proposed method showed a slight increase of 0.19% compared to baseline YOLOv7. The experimental results demonstrated that the proposed method is highly effective for improving detection accuracy without increasing computational cost.