The global navigation satellite system (GNSS) and inertial navigation system (INS) integrated navigation can realize accurate and reliable positioning for land vehicles. However, during GNSS outages, the cumulative errors caused by inertial sensors pose a threat to the GNSS/INS integrated system, especially when low-cost microelectromechanical system (MEMS) inertial sensors are utilized. The existing methods usually apply nonholonomic constraint (NHC) and machine-learning (ML)-/deep-learning (DL)-based modeling techniques to constrain the errors. Nevertheless, the performance of using NHC alone is limited if the forward velocity is not accurate enough, so additional constraints are needed, while the existing ML-/DL-based modeling techniques only utilize acceleration or angular velocity features, which are still susceptible to the stochastic errors of MEMS sensors. In this article, an ML-based modeling technique combining acceleration and frequency modulation (FM) radio signal is applied to further improve the performance of NHC-constrained MEMS-INS/GNSS integrated system. First, we derive the relationship between the received signal strength indicator (RSSI) of FM signal and vehicle distance increment to provide theoretical basis for the proposed algorithm. Then, FM signal features related to distance increment are extracted. Afterward, an availability assessment strategy is proposed to eliminate the moments when errors of using RSSI are large. Subsequently, we apply support vector regression (SVR) to estimate distance increment by combining FM signal and acceleration features. Finally, extended Kalman filter (EKF) is applied to fuse the predicted distance with INS during GNSS outages. Results show that the introduction of FM signal can significantly reduce distance estimation errors, thus improving positioning performance.