Object Detection Algorithm Based on Global Information Fusion

Abstract

The output channel of the Fully Convolutional One-Stage Object Detection (FCOS) feature extraction network drops sharply before input FPN. Feature information is severely lost. This paper proposes an object detection algorithm based on global information fusion. First, a multi-scale global aggregation module is designed. This module extracts the information of multi-scale receptive fields, aggregates global features, and enhances local features. The last layer features of the feature backbone network are downsampled, and fused with feature downsampling enhanced using a multi-scale global aggregation module. The enhanced feature upsampling and shallow feature fusion is output through FPN. The proposed algorithm uses Generalized Intersection over Union loss instead of Intersection over Union loss, which can make the target localization more accurate. The detection accuracy of the algorithm in this paper on the PASCAL VOC dataset reaches 82.8%, which is 2.0% higher than FCOS. The accuracy on the KITTI dataset reaches 82.4%, which is 4.2% higher than FCOS. And the detection speed meets the real-time requirements.