Abstract: Car accident prediction is an important problem to study for avoiding the accidents. Previous studies make the prediction for a car based either on macro factors such as geography, environment and traffic or on micro factors such as car and driver behaviors. There is rarely a study combining the two types of factors because it is difficult to collect the two types of data at the same time. However, car accidents usually result from both of the two types of factors. In addition, the current researches predict whether an accident will happen or not. There is rarely a study providing a more accurate accident probability because there is no probability label for use in the collected data. However, such a probability is useful to notify the driver in different warning levels. The OSU(Ohio State University) accident dataset of macro factors published in 2019 has some identical characteristics with the FARS(fatality analysis reporting system) dataset of macro factors and SHRP2(strategic highway research program 2) dataset of micro factors, and thus provides an opportunity to fuse them. Therefore in this paper, we obtain a dataset of both macro and micro factors. In the dataset, accident data (positive data) is fused from the OSU and FARS datasets, as well as Sim-SHRP2(simulated strategic highway research program 2) similar to the SHRP2 dataset, while safe-driving data (negative data) is obtained by ourselves driving a car. In addition, since the obtained dataset does not have any probability label, we also design a probability-level unsupervised deep learning framework to predict the accurate probability. The framework iteratively generates accurate probabilities from the obtained dataset, and is trained with the generated probabilities. The experimental results indicate our framework can predict car accidents with the obtained dataset sensitively and accurately.