Fatigue driving represents a major cause of road traffic accidents, with particularly severe consequences in passenger transport. Developing accurate fatigue detection methods resilient to in-vehicle environmental interference is crucial for enhancing road safety. This study proposes a multi-feature fatigue detection framework integrating temporal information based on YOLOv8n-SimAM for bus drivers. The SimAM attention mechanism enhances feature extraction from ocular and oral regions while suppressing cabin background interference. By incorporating the driver’s eye and mouth states with temporal information, we established a multi-feature fatigue reasoning mechanism. Comparative experiments demonstrate that YOLOv8n-SimAM achieves superior performance in precision, recall, and P_mA50-95 (mean average precision across IoU thresholds 0.5–0.95) over alternative models. Our approach achieves enhancements of 2.8% in precision, 5.4% in recall, and 7.5% in F_1-score (harmonic mean of precision and recall) over the suboptimal benchmark, without a substantial increase in model parameters. Ablation studies confirm that integrating SimAM reduces false positive rate by 48%, while the temporal fatigue inference module decreases missed detection rate by 72%. The integrated model attains an F_1-score of 0.906, an accuracy of 0.9, with corresponding missed detection and false positive rates of 0.111 and 0.087, respectively. This methodology enables reliable facial feature capture and fatigue assessment in challenging in-cabin environments, providing an effective solution for off-site monitoring of bus driver fatigue.

To further improve the regulatory efficiency of the driver training industry and promote the development of “Internet + supervision” in the driver training industry, an off-site supervision method for the driver training industry based on geofence technology and geospatial analysis methods is studied. This method aims to automatically identify and comprehensively supervise whether training vehicles operate in accordance with specified routes and times. Through spatiotemporal matching and spatial mapping of multi-source heterogeneous data such as trajectory data of training vehicles from driving training institutions and geofence data, a multi-source dataset for industry supervision is established. Using the Shapely geospatial analysis library, based on the DE-9IM model, and combined with the multi-source data infrastructure, real-time supervision of training vehicles and automatic identification of violations are realized. The results show that the off-site supervision method proposed in this study can achieve precise supervision of the driving training industry, with a supervision accuracy rate as high as 99.87%. The identification results can serve as an important basis for relevant industry regulatory and law enforcement departments to carry out off-site supervision and early warning in the industry, and promote the intelligent transformation of off-site supervision in the driving training industry.