Authors

Abstract

Accurate detection of multiple behaviors in group-housed pigs was important for precision livestock farming and intelligent farm management. This study proposed RAFE-DETR, an improved detector based on RT-DETR, for recognizing standing, lying, feeding, drinking, and fighting in overhead surveillance images. RFAConv was embedded into RepViT blocks to construct the RFA-RepViT backbone for stronger local feature extraction. The original intra-scale interaction module was replaced with BiFormer to improve contextual modeling. The neck was redesigned with ASF-CSA to enhance adaptive multi-scale fusion, and Focaler-Shape-IoU was introduced to refine box regression. Experiments were conducted on a five-class dataset reconstructed from public surveillance videos. The proposed model achieved 93.9% precision, 92.7% recall, and 94.2% mean average precision at an intersection-over-union threshold of 0.5. Compared with RT-DETR-L, these values increased by 1.4, 2.8, and 3.0 percentage points, respectively. At the same time, the number of parameters decreased from 32.0 M to 21.9 M, and GFLOPs decreased from 103.5 to 77.0. Supplementary experiments on a second public dataset supported the robustness of the method. Deployment on Jetson Orin NX Super reached 13.8 and 19.1 frames per second under PyTorch and TensorRT, respectively, indicating good edge-deployment potential.

Abstract in Chinese

群养猪多行为检测是精准养殖和智能化猪场管理中的重要环节。针对俯视监控图像中的站立、躺卧、采食、饮水和打斗五类行为识别任务，本研究提出了一种基于 RT-DETR 的改进检测模型 RAFE-DETR。通过在 RepViT 模块中嵌入 RFAConv，构建了 RFA-RepViT 主干网络，以增强局部特征提取能力。原始同尺度特征交互模块被替换为 BiFormer，以提升上下文建模能力。Neck 结构被重构为 ASF-CSA，以增强自适应多尺度特征融合，并引入 Focaler-Shape-IoU 以优化边界框回归质量。实验基于由公开监控视频重建的五类行为数据集开展。结果表明，RAFE-DETR 在主数据集上的精确率、召回率和 mAP@0.5 分别达到 93.9%、92.7% 和 94.2%。与 RT-DETR-L 相比，这三项指标分别提高了 1.4、2.8 和 3.0 个百分点。同时，参数量由 32.0 M 降至 21.9 M，GFLOPs 由 103.5 降至 77.0。第二公共数据集上的补充实验进一步验证了该方法的稳定性。部署结果表明，模型在 Jetson Orin NX Super 平台上采用 PyTorch 和 TensorRT 后端时，推理速度分别达到 13.8 FPS 和 19.1 FPS，表明该模型具有较好的边缘部署潜力。