DED-YOLOv8:Dense pedestrian detection algorithm based on YOLOv8
Xiao Lin Li, Lishun Ma · 2024
Aiming at the problem that pedestrian occlusion and scale changes in dense scenes make the model prone to missed detection and false detection, a dense pedestrian detection algorithm improved from YOLOv8n is proposed: DED-YOLOv8. Firstly, the backbone network improved by Deformable ConvNets_V2 (Dcnv2) is employed to boost the capacity of extracting pedestrian features. Then, the Efficient Multi-Scale Attention (EMA) module is introduced into the baseline model. Finally, The Dynamic Head (Dyhead) is brought in to heighten the attention towards multi-scale pedestrian features and strengthen the expression capacity of the detection head. Through the analysis of experimental results, on the WiderPerson public dense pedestrian dataset, in comparison with the model of YOLOv8n, the [email protected] and [email protected]:0.95 of DED-YOLOv8 have risen respectively by 2.5% and 1.7%. It has been greatly improved, and it also has shown strong competitiveness when compared with other advanced human detection models.