Aiming at improving the prediction and judgment of rollover risk in rollover warning， a more efficient and accurate rollover warning algorithm was proposed to provide an important basis for drivers or other driving assistance systems to determine the intervention time of vehicle control. Firstly， a 3-DOF vehicle pre-warning reference model was established. The phase-plane method was selected to divide the roll stability region as the rollover index， an improved time-to-rollover （TTR） algorithm was designed and TTR was calculated according to the response of the 3-DOF vehicle model. The analysis results show that the phase plane rollover index is close to the actual lateral-load transfer rate （LTR）， which is more accurate than the common expression of LTR， and the improved TTR is closer to the actual TTR. Then， to improve the computational efficiency of pre-warning， a long short-term memory （LSTM） model was established to replace the improved TTR algorithm and the TTR value output by the model was used as the basis for vehicle pre-warning control. Finally， the LSTM model was trained by collecting data through driver-in-the-loop （DIL） tests. In two working conditions， the proposed rollover warning method was verified to have the accuracy of rollover risk prediction and higher real-time performance.