Small-scale Pedestrian Detection Based on Feature Enhancement Strategy

In pedestrian detection, small-scale pedestrians are often missed and mistakenly detected. In order to improve detection precision and reduce miss detection rate of small-scale pedestrians, a feature enhancement module is proposed. First, considering the problem that small-scale pedestrians feature gradually decreases as network goes deeper, feature fusion strategy breaks through the constraints of feature pyramid structure and fuses deep and shallow feature maps to retain lots of small-scale pedestrian features. Then, considering the problem that small-scale pedestrian features are easily confused with background information, self-attention module combined with channel attention module models the spatial and channel correlation of feature maps, using small-scale pedestrian contextual information and channel information to enhance small-scale pedestrian features and suppress background information. Finally, a small-scale pedestrian detector is constructed based on the feature enhancement module. For small-scale pedestrians, the proposed algorithm has 19.8% detection accuracy, 22 frames per second speed on CrowdHuman dataset and 13.1% miss rate on CityPersons dataset. The results show that the proposed algorithm performs better than other compared algorithms for small-scale pedestrian detection and achieves faster detection speed.