For conventional particle filter-based vehicle speed estimators, the estimation performance deteriorates if the proposal distribution is inconsistent with the actual distribution. In this paper, an improved particle filter is proposed to tackle this problem by modifying the proposal distribution. Firstly, based on vehicle kinematics and sensor characteristics, the state transition equation and the observation equation are established. Secondly, the error between the sensor measurement and the particle state is employed to design the proposal distribution correction term. Then, the process noise in the state transition equation is made adaptive to enhance robustness. Lastly, CarSim-Simulink co-simulation is conducted to compare the proposed speed estimator and the basic particle filter, under the double lane change maneuver and the sine wave steer input maneuver. For the former maneuver, using the proposed estimator, the average absolute error (ATE) of the estimated longitudinal velocity is reduced by 40.25%, and that of the estimated lateral velocity is decreased by 55.71%. For the latter maneuver, the ATE of the estimated longitudinal velocity is reduced by 47.00%, and that of the estimated lateral velocity is decreased by 41.21%.