Multi-objective real-time detection method for construction machinery unloading scene based on improved YOLOv10

With the rapid development of infrastructure, in the intelligent cluster working of construction machinery, especially loaders, excavators and trucks interact frequently with each other, and there are often overlapping targets and dust occlusion, so the ability to face the complex scene environment perception needs to be improved. This study proposes a multi-target detection method. The method addresses low detection accuracy caused by dense occlusion. It also reduces the high computation cost in localized construction machinery unloading scenes. First, the Omni-Dimensional Dynamic Convolution is introduced into the YOLOv10n algorithm to improve the flexibility in complex multi-target scenarios. Second, the Online Convolutional Re-parameterization is added to improve the model learning ability without increasing the inference cost. Finally, a dataset of construction machinery working scenarios is constructed and annotated to train and test the model. The results show that the improved YOLOv10n algorithm can achieve 96.8% multi-target detection accuracy for localized scenes of construction machinery unloading working, and the detection speed is 60.37 FPS, which is improved by 8.2% and 9.43 FPS compared with the original YOLOv10n detection network, and meets the real-time requirements. The method provides a new idea for target detection of complex construction machinery, which helps to promote the development of construction machinery in the direction of clustering and intelligent working.