Automated Construction Site Monitoring Based on Improved YOLOv8-seg Instance Segmentation Algorithm

Utilizing Unmanned Aerial Vehicles (UAV) and instance segmentation for construction site monitoring(such as construction machinery and operation surfaces) offers a significant leap in management efficiency over traditional manual supervision methods. However, in UAV-based remote sensing images, the subtle presence of construction machinery and the image features resemblances among various operational surfaces make it difficult to segment instances. To address these challenges, this study proposed a novel instance segmentation model based on the YOLOv8-seg model. Given the unique challenges, the proposed model makes three improvements to the original YOLOv8-seg model. First, the paper incorporates the FocalNext module, which extends the sense field of the convolutional kernel to capture contextual data and integrates multilevel features, enhancing the perception of local details. Second, the paper incorporates the Efficient Multiscale Attention (EMA) module, which refines image features by emphasizing spatial-channel interactions and adeptly contrasts patterns across scales to detect nuances overlooked by conventional models, aiding in distinguishing similar construction operation surfaces. Last, given the intricate nature of construction site images, this paper incorporates the Context Aggregation module, which enhances pixel analysis by intelligently modulating feature weights to highlight essential global contexts. The ablation experiment demonstrates that the enhancements perform well on the YOLOv8-seg two variants model. Comparative experimental results show that the improved model significantly outperforms existing instance segmentation models regarding model performance, complexity, and inference speed. Overall, the improved YOLOv8-seg model balances model performance and computational complexity to meet the needs of edge device deployment in field monitoring.