YOLOv8-GDI: A lightweight YOLOv8 for real-time Pedestrian Detection

Aiming at the current challenges of low detection efficiency in real-time pedestrian detection algorithms, complex model structures, and difficulties in implementing them on mobile devices, we proposed YOLOv8-GDI, a lightweight pedestrian detection model designed based on the backbone of YOLOv8n. Firstly, the C2f-Ghost module is constructed by incorporating the Ghost Convolution in the model, in order to decrease the number of floating-point operations during feature channel fusion and enhance feature expression capability. In addition, Dynamic Task Align Detection Head(DTADH) was designed to facilitate the interaction between the model classification and localization to further enhance the model's detection performance. Finally, the Inner-CIoU loss function was introduced into the YOLOv8n to improve the ability to localize features of different sizes. The experimental results show that compared with YOLOv8n, YOLOv8-GDI reduced the number of parameters by 50%, decreased the computational complexity by 25%, and maintained a high detection accuracy on the Wider person dataset. This suggested that the improved model is more lightweight and has better application value in the direction of real-time pedestrian detection.