With the rapid development of deep learning technology, object detection has been widely applied in multiple fields. However, small object detection has limited detection performance due to its small size and unclear features. To address this issue, this paper proposes an improved object detection model based on YOLOv8. This model integrates optimization strategies such as ghost bottleneck network, multi-scale free attention module, improved feature pyramid network, and dynamic Soft NMS, aiming to improve the detection accuracy of dense small targets and the computational efficiency of the model. Through experimental validation on a self-made dataset, it has been demonstrated that the improved YOLO model outperforms existing mainstream models in terms of precision, recall rate, and mAP@0.5, which are key metrics, effectively balancing the model's parameter count and floating-point computational load. The experimental results show that the proposed method achieves model lightweighting while ensuring detection accuracy, providing an effective solution for object detection tasks on resource limited embedded devices.