Lightweight Rotated Object Detection Model for Pallets
Minglin Sun, Aidong Ge, Mingcan Sun · 2025
To address the issues of excessive computational overhead, oversized parameter volume, and slow deployment efficiency of standard rotated pallet detection models on intelligent forklifts, we propose a lightweight rotated object detection framework based on YOLOv11n. Firstly, the backbone of YOLOv11n is replaced with StarNet to reduce unnecessary computations and improve computational efficiency. Furthermore, HSFPN is used to construct the Neck, reducing parameter count and floating-point operations (GFLOPs). Lastly, the head of YOLOv11n is improved using a Lightweight Asymmetric Detection Head, which reduces model parameters and enhances detection speed. Experiments were conducted on a selfconstructed standard pallet dataset. Compared to YOLOv11n, the improved model achieves a 61.93% reduction in parameters, a 54.10% compression in model size, a 50% decrease in GFLOPs and a 25% increase in FPS, without sacrificing detection accuracy. The lightweight model saves a considerable amount of computing power for the computing unit of intelligent forklifts, enabling better handling of other tasks and improving overall operational speed.