Small-scale Pedestrian Detection Method based on Attention Mechanism and Adaptive Feature Fusion

Abstract

In pedestrian detection, high detection accuracy has been achieved for large-scale pedestrians, but for small-scale pedestrians, the detection effect needs to be further improved. In order to improve the detection accuracy of small-scale pedestrians, this paper proposes a YOLOv4 small-scale pedestrian detection algorithm based on the fusion of attention and weighted features. In order to enhance effective features and suppress ineffective features, an attention mechanism that adapts features is introduced in the backbone network CSPdarknet53 to reduce the interference to small-scale pedestrian detection. In order to better integrate features with inconsistent semantics and scales, adaptive channel-weighted feature fusion is used in the feature pyramid, so that the deep and shallow features focus on pedestrian targets of corresponding scales. The logarithmic mean missed detection rate (LAMR) on the small-scale pedestrian test sets “Far” and “Medium” of the Caltech public data set has decreased: 9.67% And 7.38%. Compared with other pedestrian detection algorithms, it has obvious advantages for small-scale pedestrian detection.