BFNet: A Lightweight Network Designed for Pedestrian Detection

Many advanced pedestrian detection models have achieved incredible results. However, these high-accuracy models have huge network scales and heavy inference computation loads, which may lead to the failure in high real-time industrial fields. In this paper, we propose a Branch Fusion Network (BFNet) to reduce inference computations and improve the inference speed with the high detection accuracy. The inference computations of most pedestrian detection models mainly concentrate upon the feature extraction stage. The proposed BFNet shunts the features in the feature extraction stage on channel dimension through the branch structure. Only part of the features passes through the original model branch with heavy computation loads, while the rest of the features pass through the lightweight branch. Moreover, multi-branch features inside BFNet are fused to make feature information complementary and enhance the learning ability of the network. The proposed networks focus on both real-time performance and detection accuracy. The experimental results on pedestrian datasets indicate that the proposed two-branch BFNet sacrifices only a small amount of detection accuracy but reduces the computations by 46.3%.