This paper proposes a low-cost, real-time, and robust indoor autonomous localization algorithm that achieves high-precision state estimation by fusing ultra-wideband (UWB) and inertial measurement units (IMU). A multi-source data fusion algorithm based on the error-state Kalman filter is adopted to effectively suppress the state estimation error caused by UWB measurement noise. Aiming at the non-line-of-sight (NLOS) problem caused by occlusion that may occur in UWB systems, an adaptive anomaly detection method is developed. In this method, the Isolation Forest (IForest) is used to evaluate continuous UWB data to update the anomaly threshold, with the mean of normal data serving as the anomaly threshold. The adaptive anomaly threshold is used to filter new data, adjusting the confidence level before data fusion to mitigate the impact of anomalies. Moreover, to efficiently utilize UWB data, data from all anchors in a single observation is employed as an observation set to update the state of vehicles. The experiments conducted on public datasets and in indoor environments show that the proposed algorithm can achieve a localization accuracy of approximately 0.1 m, representing a 75.0% to 87.9% improvement over the multilateration method.