Lightweight household object detection algorithm based on GFM-YOLOv8

To address the challenges of accurate detection in home environments where object scales vary significantly and backgrounds are complex, this study proposes a lightweight household object detection model named GFM-YOLOv8 based on improved YOLOv8. Firstly, use GhostC2f instead of C2f in the backbone and neck networks of the model, and replace some Conv layers in the network with GhostConv to reduce the complexity and parameter count of the model. Secondly, a Focal modulation network is integrated to replace the original SPPF module, enhancing feature fusion capabilities and enriching semantic information during feature extraction. Finally, an MPDIoU loss function with dynamic feature adaptation mechanism is adopted to better focus on bounding box shape and scale, enabling more accurate bounding box regression. Experimental results demonstrate that the optimized model achieves a mean detection accuracy of 77.93%, representing a 2.97-percentage-point improvement over the original model. Meanwhile, computational cost and parameter count are reduced by 39.02% and 40.20% respectively. When deployed on the Jetson Nano 4GB edge computing device, the model achieves a detection frame rate of 47.3 frames per second. This approach effectively solves intelligent detection of household objects in complex backgrounds, meets deployment requirements for low-power embedded devices, and contributes to advancing intelligent service technologies for home robots.