针对堆石混凝土坝施工进度监控需求,围绕堆石入仓工序提出了一种自卸汽车作业状态智能监测方法。采用改进的YOLO11-Pose模型识别自卸汽车并定位车体前、中、后3个关键点,结合关键点归一化坐标和车体轴向投影速度等时序运动学特征,构建车体运动状态与车斗动作状态的并行判别方法,实现倒车入仓、停车静止、行进出仓及车斗抬升、复原等状态识别。结果表明,模型在关键点定位和目标检测任务上的平均精度指标mAP_50-95分别达到0.98和0.75。7 103帧有效时序图像的验证结果显示,该状态判别方法对车辆运动状态和车斗动作状态的识别准确率分别达到90%和96%,联合判别准确率为93%;各类状态对应的F1值分别不低于0.89和0.96。该方法可部署于堆石混凝土坝施工仓面监控场景,有效实现自卸汽车作业状态的自动判别,为堆石入仓工序智能化监测提供技术支撑。

[Objective] Rockfill placement is a key procedure in rock-filled concrete dam construction. As the main equipment for transporting rock materials, dump trucks reflect unloading coordination and construction progress through their working states. Existing studies have formed a technical chain from object detection and keypoint localization to behavior recognition, with preliminary applications in dam-site vehicle unloading recognition. However, recognition based only on unloading/non-unloading or fine segmentation limited to unloading, cannot fully meet the monitoring needs of the complete “entering the bay-unloading-leaving the bay” process. This study investigates intelligent monitoring of dump truck working states in rockfill placement. By combining keypoint perception with continuous-frame temporal kinematic analysis, the proposed method automatically identifies vehicle body motion and dump body action states, aiming to support real-time operation monitoring and dynamic optimization of transportation organization. [Methods] YOLO11-POSE was used to detect dump trucks in construction videos and extract three keypoints at the front, middle, and rear of the vehicle body. The front and rear keypoints describe overall vehicle motion, while the middle keypoint reflects dump body height variation. To improve multi-scale representation under surveillance views, RepLKBlock and BiFPN-Concat were introduced into YOLO11-POSE to build an enhanced feature fusion network. Based on continuous-frame keypoint sequences, the front-rear keypoint distance was used as the scale reference for coordinate normalization, and temporal kinematic parameters such as vehicle axial projection velocity and dump body height variation rate were constructed. Since vehicle motion and dump body lifting/lowering are related but partly independent, a two-dimensional parallel classification strategy was adopted. The working state was decomposed into vehicle body action and dump body action.[Results] The improved YOLO11-POSE model effectively detects dump trucks and locates the three keypoints, achieving mAP_50-95 values of 0.98 and 0.75 for keypoint localization and object detection, respectively. Considering accuracy, parameter efficiency, and training stability, the model with RepLKBlock and BiFPN-Concat was selected for state recognition. Validation on 7,103 effective temporal frames showed that the vehicle motion classifier achieved 90% accuracy, while the dump body classifier reached 96%. The F1 scores of all vehicle states were no lower than 0.89, and those of all dump body states were no lower than 0.96. The joint accuracy of the two classifiers reached 93%. These results demonstrate that the method can identify key states, including reversing into the bay, stationary parking, driving out of the bay, dump body lifting, resetting, and no action. Recognition errors mainly occurred near state transitions, where short-interval motion parameters changed only slightly, reducing classification stability. The relatively small sample proportion,such as reversing-into-bay states, also limited feature learning. Illumination changes and viewpoint disturbances caused a few additional misclassifications. Engineering application results show that the overlaid state labels reflect key dynamic parameters and working states in real time, generally matching visual observations. [Conclusion] Compared with single-frame detection or single sequential state description methods, the proposed method separately models vehicle body actions and dump body actions, better representing their temporal combinations and improving the specificity and physical interpretability of state description. Validation and engineering application results confirm that the method effectively identifies dump truck working states and provides a feasible approach for intelligent video monitoring in rock-filled concrete construction. Future work will optimize temporal feature representation, improve recognition of complex transitions, and expand validation across different environments and vehicle types.