Multi-target tracking is the basis to ensure the safety and efficiency of autonomous driving, and the obtained data are widely used in the upper applications of autonomous driving, such as motion planning and driving decision making. The traditional multi-target tracking method often has the phenomenon of target loss and dislocation in the occlusion environment. To solve the problem, a robust tracking method based on heterogeneous radar fusion and occlusion prediction model is proposed. Firstly, based on the local observation consistency equation of lidar and millimeter-wave radar, a multi-sensor dynamic self-calibration algorithm based on multi-target motion constraint and global maximum matching is proposed. Secondly, a hybrid supervised target position prediction method based on heterogeneous radar fusion unscented Kalman filter and long and short time series neural network is proposed to solve the problem of tracking interruption caused by missing observation data in complete occlusion environment. Experiments show that the proposed method can effectively complete at least 81% of the broken multi-vehicle target tracks in the complete occlusion environment, which can achieve more reliable multi-vehicle target tracking compared with the most advanced methods.