改进RT-DETR的油田人员异常行为检测OA

To address the frequent safety accidents caused by non-standard operations at oilfield worksites,this study proposes an efficient improved RT-DETR algorithm,named HCH-DETR,for ab-normal behavior detection in complex oilfield scenarios.Firstly,a novel backbone network is designed by integrating the dual-branch High-Frequency Enhancement Residual Block(HFERB)and the Cross Stage Partial(CSP)structure.This design enhances the model's ability to extract high-frequency detailed features while effectively reducing the model's computational complexity.Secondly,to tackle the challen-ges such as large variations in target scale and complex backgrounds in oilfield monitoring,a Context-Guided Feature Reconstruction Feature Pyramid Network(CGFRPN)is proposed.It employs Rectangular Self-Calibration Attention(RCA)to strengthen multi-scale feature fusion,improving the detection accu-racy for multi-scale targets and bolstering robustness in complex scenes.Finally,the Haar Wavelet Down-sampling(HWD)module is introduced to optimize traditional downsampling,thereby improving the mod-el's capability for small-target detection.Experimental validation on a self-constructed oilfield dataset shows that the proposed model achieves 85.4%mAP@0.5 and 55.1%mAP@0.5:0.95,representing im-provements of 3.2 percentage points and 2.4 percentage points respectively over the original RT-DETR model.Meanwhile,the computational complexity is reduced by 7.1×109,and the number of parameters is decreased by 6.5×106.Ablation experiments verify the effectiveness of each improved module,and generalization experiments demonstrate that the model also achieves improved accuracy on the VisDrone dataset.