Accurate collision prediction algorithms are important to provide drivers reliable warning messages. This study introduces a novel approach for collision prediction at intersections. The algorithm involves the use of an index called `minimal future distance (MFD)', which is defined to be a future distance (FD) between the subject vehicle and the primary other vehicle, and a two-level dynamic threshold for performing the collision prediction task. Real-time vehicle motion information and surrounding road geometry are utilised to forecast FDs and identify MFD within an upcoming time horizon. The dynamic threshold in both emergency warning and normal warning situations consists of two parts, a vehicle heading direction-related part and a speed value-related part. Potential collisions are determined by the comparison of MFD with different dynamic thresholds in different driving scenarios. The combined use of vehicle real-time state and road geometry in the algorithm significantly increased the collision prediction accuracy. Furthermore, the use of dynamic thresholds ensured the promptness and robustness of the collision warning system. Simulation results show that the false positive rate and false negative rate of severe collisions at intersections can be robustly eliminated and those of marginal collisions can be kept low at 2.4 and 3.6%, respectively.