This paper investigates a cascaded robust Kalman filter framework for cooperative localization of vehicles subject to unreliable information from onboard sensors and vehicular ad hoc networks (VANETs). Specially, an adaptive-threshold outlier-robust local filter is first introduced that using a machine learning-based dynamic parameter selection approach. Then, a three-stage weighting function, and K-means clustering to classify residuals into reliable and outlier sets, with the maximum residual from the reliable set used to adaptively determine the threshold. Finally, by employing a dual-statistics joint test to extract two-dimensional statistical features from residual matrices, an unreliable information isolation algorithm with low computational complexity is developed. Experimental results demonstrate that the method outperforms the state-of-the-art results with ensured accuracy and robustness for cooperative vehicle localization in urban environments.