EdgeYOLO-Light: A Lightweight YOLOv5 Object Detection Model for Embedded CPU Platforms

To address the challenges of high inference latency, large parameter size, and poor performance in small object detection when using traditional YOLO models on embedded CPU platforms, we propose a lightweight YOLO model named EdgeYOLO-Light. Based on YOLOv5s, this model introduces optimizations in architecture, upsampling strategies, and the loss function. First, a more efficient and compact FasterNet structure is integrated into the backbone to reduce computational complexity and accelerate feature extraction. Second, a DySample module is introduced in the upsampling phase to improve feature restoration quality and efficiency while reducing parameters and computational overhead. Additionally, we adopt a Focal-SIoU loss function that combines the Focal mechanism with SIoU to handle sample imbalance and the difficulty of fitting small objects, enhancing detection accuracy in complex scenes. To further compress the model size, an improved C3_Faster module is adopted in the neck, maintaining feature extraction ability while minimizing computation. Experiments on the VisDrone2019 dataset show that EdgeYOLO-Light reduces inference time by 11% and improves mAP by 5 % compared to YOLOv5s on CPU. Ablation studies confirm the effectiveness of each module, demonstrating a good balance between model lightweightness and accuracy. This method offers a promising solution for efficient and accurate object detection in resource-constrained embedded platforms.