In current civil aircraft cockpit flight tasks, primary flight display (PFD) state information is still mainly interpreted manually. Owing to the dense interface layout, diverse display forms, and limited accessibility of airborne data interfaces caused by permission management, airworthiness certification, and platform openness, external-camera-based non-contact visual perception provides a flexible solution for PFD state acquisition. This article proposes a vision-based method for PFD multi-state extraction and temporal correction to provide state inputs for action triggering and auxiliary decision-making during flight tasks. An improved YOLOv8 model is used for PFD region detection, followed by ORB-based perspective correction to improve the spatial consistency of state parsing. According to the visual features, color attributes, and geometric structures of different display elements, corresponding parsing strategies are designed to extract multiple key state variables. A second-order constant-velocity state-space model is further established, in which sequential Kalman filtering and residual-statistics-based anomaly detection are combined for recursive estimation and correction. Experiments conducted in an X-Plane-based semi-physical cockpit platform show that the proposed method achieves stable PFD localization, with MCE, MWE, and MHE of 5.44, 14.15, and 5.81 pixels, respectively. For continuous-state estimation, the MAE/RMSE values of airspeed, climb rate, barometric altitude, and radio altitude are 0.06/0.08 kt, 0.07/0.14, 10.16/25.98 ft, and 1.55/1.98 ft, respectively. The results show that the proposed temporal correction method can suppress abnormal observations and improve the temporal stability of PFD multi-state parsing under controlled semi-physical simulation conditions.