Multi-layer Feature Fusion Network with Atrous Convolution for Pedestrian Detection

In this paper, we present a simple but effective framework K-AFPN that incorporates feature pyramid method for small-size pedestrian detection, fully utilizing the lower-layer detail features and higher-layer semantic features. The method not only enhances the robustness of the features, but also improves the discrimination of the feature maps, achieving competitive accuracy. In addition, atrous convolution is used to optimize the network for high-resolution feature maps, avoiding information loss caused by frequent down or up sampling. On top of the backbone network, K-means algorithm is used to obtain optimal initial anchor base sizes, which reduces computational costs and improves location accuracy. Hence, our method pays more concentration on pedestrians, especially those of relatively small size. Comprehensive experimental results on two classic pedestrian benchmarks illustrate the effectiveness of the proposed approach.