基于MBD-RTDETR的多场景生猪计数技术OA

To facilitate the digital management of swine assets,an improved algorithm based on the MBD-RTDETR model is proposed for counting pigs across various scenarios.Built upon the RTDETR framework,this algorithm incorporates the MobileNetV4 backbone network for optimization,thereby reducing model complexity and parameter count while enhancing operational efficiency and deployment flexibility.Furthermore,the introduction of a BiFPN structure strengthens the model's capability to capture multi-scale features,while the integration of a DAttention mechanism improves overall accuracy.To further improve performance,the algorithm also combines masking and ByteTrack tracking algorithms to ultimately achieve the identification of the number of pigs inside the pen and through the passageway.Compared to the baseline RTDETR model,MBD-RTDETR demonstrates improvements across multiple performance metrics:parameter count is reduced by 48.72%,computational load by 36.03%,and model size by 48.31%;concurrently,the P(Precision)increases by 1.45 percentage points,the F1 score improves by 0.42,and the AP(Average Precision)rises by 1.07 percentage points.The improved MBD-RTDETR model serves as a valuable technical reference for smart swine farming applications.