PFF-FPN: A Parallel Feature Fusion Module Based on FPN in Pedestrian Detection

Abstract

Feature extraction in pedestrian detection is a challenging problem due to the different sizes of pedestrians and occlusion in pedestrians. Currently, Feature Pypyramid Networks(FPN) structure is usually used in pedestrian detection networks for feature extraction but aiming at the characteristics of pedestrian detection tasks it may not be effective in extracting important layer feature information. Therefore, this paper proposes a module based on PFN structure with parallel feature fusion named PFF-FPN. PFF-FPN uses three different FPNs to extract feature and fuses the corresponding layer feature to reinforce the focused layer feature information. In pedestrian detection task PFF-FPN can be adapted to different network frameworks and it also gets a good performance.