Pedestrian detection algorithm based on multi-scale feature extraction and attention feature fusion

Abstract

Despite rapid development in anchor-free pedestrian detection algorithms, an optimistic trade-off between detection accuracy and efficiency is still far from being achieved. In this study, we proposed a new pedestrian detection algorithm based on multi-scale feature extraction and attention feature fusion, which is called MSAF-Net. Firstly, we designed a multi-scale dilate residual module to expand receptive fields while maintaining feature map sizes and improving the spatial sensitivity of information features. Secondly, through a joint attention feature fusion mechanism, the interaction of channel and space joint information features was captured. The context and initial features were integrated into strengthen extracted image features. Finally, a channel attention guidance mask branch was added to a detector to locate pedestrian position information accurately, and to improve the model's robustness. The experimental results demonstrated that our algorithm achieved satisfactory results on two well established datasets, i.e., CityPersons and Caltech. For CityPersons, the values of MR−2 were 9.04% and 40.41% under reasonable and heavy occlusion conditions, which are 0.36% and 3.89% better than the suboptimal comparison detection methods, proving the effectiveness and advancement of the proposed algorithm in this study.