Towards Better Pedestrian Detection Using Multi-Scale CSPN and Dual Attention

Pedestrian detection is a significant research direction in the computer vision, but the detection performance of existing pedestrian detection algorithms is inadequate. Therefore, this article proposes a novel algorithm to improve the anchor-free pedestrian detection algorithm. First, the multi-scale CSPN module is used to deepen the network depth, further extract semantic information on multiple scales, and improve detection performance. Moreover, the dual attention module based on feature fusion is used to effectively fuse features of different scales, assigning new weights to the fused features in the two dimensions of space and channel. Experiments show our method reduces MR−2 by 0.10%, 2.60% and 0.98% on the Reasonable, Heavy Occlusion and ALL of the Caltech pedestrian dataset, which is better than the existing algorithms.