An improved object detection algorithm was proposed aiming at the issue of insufficient recognition accuracy of the original YOLOv8n algorithm in the task of drone-based warehouse inventory. Inner-EIoU, which was adaptable to target overlap and relaxed aspect ratio constraints, was introduced as the bounding box loss function in order to enhance the recognition capability for densely stacked goods. The standard convolution in the C2f module was replaced with omni-dimensional dynamic convolution that incorporated a multi-dimensional attention mechanism in the feature extraction stage. The FocalModulation method was applied to the spatial pyramid pooling-fast module in order to improve feature representation. A channel pruning method based on the LAMP criterion was used for model lightweighting in order to reduce the parameter scale of the model and increase inference speed. The experimental results on a self-built warehouse dataset showed that the improved algorithm achieved mAP50 of 94.2%, with model size of only 2.8 MB and inference speed of 135 frames/s, demonstrating the feasibility for deployment on edge computing device.