DMFFNet: Dual-Mode Multi-Scale Feature Fusion-Based Pedestrian Detection Method

Most contemporary pedestrian detection algorithms are based on visible light image detection. However, in environments with dim light, small targets, and easily occluded cluttered backgrounds, single-mode images relying color, texture, other features cannot adequately represent the feature information of targets; as a result, large number targets lost algorithm performance is not good. To address this problem, we propose dual-modal multi-scale fusion network (DMFFNet). First, use MobileNet v3 backbone to extract input for attention (MFA) module, combining idea mechanism. Second, deeply fuse output by MFA double deep (DDFF) module enhance semantic geometric target. Finally, optimize loss function reflect distance between predicted box real more realistically well ability toward predicting difficult samples. We performed multi-directional evaluations KAIST dual-light dataset visible-thermal infrared (VTI) our laboratory through comparative ablation experiments. The overall MR-2 9.26%, partial occlusion, severe occlusion 5.17%, 23.35%, 47.31%, respectively. VIT results show that performs detection, especially when target was occluded.