Precise state-of-charge (SoC) estimation in lithium-ion batteries is vital for battery management systems (BMS) in electric vehicles. Under low-temperature conditions, increased electrochemical polarization, internal resistance, and nonlinear OCV-SoC dynamics impair traditional Extended Kalman Filter (EKF) and standard Particle Filter (PF) methods. This study introduces an Extended Kalman Particle Filter (EKPF) in which each particle maintains its own EKF covariance matrix and completes a full prediction-update cycle at each step, generating a measurement-informed Gaussian proposal distribution that simultaneously mitigates linearization errors and particle degeneracy. The estimator employs a 1RC equivalent circuit model parameterized by Hybrid Pulse Power Characterization tests. Validation used data from an LG 18650HG2 NMC cell across four drive cycles (US06, UDDS, MIX1, MIX2) and six temperatures (−20 °C to 40 °C). The EKPF achieves MAE below 3% and RMSE below 4% at −20 °C, reducing MAE by 82–95% and RMSE by 83–94% relative to EKF and PF. Sensitivity analysis confirms N = 100 particles as optimal; larger ensembles cause weight collapse on flat OCV-SoC regions and degrade accuracy. Real-time deployment on an STM32H723ZG ARM Cortex-M7 microcontroller yields an average execution time of 2.252 ms and a worst-case execution time of 2.307 ms per step, within the 10 ms automotive BMS task period, with a memory footprint of 11.2 KB Flash and 10.9 KB RAM, confirming feasibility for production-grade embedded BMS deployment.