EVA-ODADD: an efficient object detection algorithm with dynamic-sampling decoupled-head for embedded vision applications

Traditional object detection algorithms deployed on embedded terminals often face a critical limitation: the conflict between high detection accuracy and restricted computational resources (e.g., storage and floating-point operations). Existing lightweight models often compromise accuracy for speed, whereas high-performance models are too heavy for practical deployment. To address this challenge, we propose EVA-ODADD, an efficient end-to-end detector based on YOLOv5. The method integrates three key innovations: (1) a cross-channel multiplicative expansion module combined with C3-MViT to enhance non-linear and global feature representation, (2) an SD-AttentionNet to aggregate multi-scale features and focus on key regions, and (3) a dynamic sampling decoupled head that improves classification and regression through fine-grained feature extraction. Experimental results on the PASCAL VOC dataset demonstrate that EVA-ODADD achieves an AP@0.5–0.95 improvement of 11.0% and an AP@0.5 increase of 6.3% compared with YOLOv5n, with only 3.49M parameters and 6.3 giga floating-point operations (GFLOPs). Moreover, evaluations on the MS COCO dataset validate the model’s generalization capability, achieving increases of 6.5% in AP@0.5–0.95 and 5.8% in AP@0.5, proving its suitability for versatile deployment on resource-constrained embedded platforms. The code is available at https://github.com/miaolinghui/EVA-ODADD.git