In this paper, we propose a method for predicting moving obstacles using a probabilistic representation of the real world. In this method, we introduce a model with the spatial and temporal correlations among points in the area. Then, we map the probabilistic information of the detected obstacles onto the model. The existence of obstacles at each location and future time slot is predicted based on the model. In addition, we also propose a method for real-time updating of the probabilistic representation through edge-cloud cooperation. In this method, we divide the target area into multiple subareas and deploy an edge computer for each subarea. Each edge computer updates the model in the corresponding area. We also deploy a cloud computing environment or a high-performance computer. The cloud aggregates model of edge computers. The cloud also shares its prediction results with the edge computers. By utilizing the prediction results from the clouds, each edge computer can predict the obstacles coming from the outside of its corresponding subarea. In this paper, we demonstrate that our approach effectively identifies locations with the risk of the existence of the obstacles through edge-cloud cooperation by simulation.