Global–Local Feature Fusion Network for Visible–Infrared Vehicle Detection

Visible-infrared vehicle target detection aims to pinpoint the location and class of vehicles by fusing the favorable complementary information of visible-infrared image pairs. However, most of detection methods cannot obtain ideal detection performance when visible-infrared image pairs are captured in low lighting environment. To solve this issue, we propose a global-local feature fusion network, which can adaptively integrate the saliency information from visible-infrared image pairs. Initially, a dual-stream ResNet-50 network is designed to extract cross-modal features from visible-infrared image pairs. Then, a global-local feature fusion module (GLF) is proposed to merge the multi-modality features. Finally, the detection head utilizes the fused features of the deep interaction to get the detection results. Experiments on the DroneVehicle and LLVIP datasets show that the proposed method is increased by 7.4% and 1.2 % compared to recently proposed methods, respectively.