The paper concerns the description of a nowcasting system used for flash floods mitigation, which is based on ground radar images acquired at local scale. The forecast of rain quantity is usually obtained either with numerical weather prediction models or with image-based methods. The first solution is suitable for temporal horizons above three hours due to the computational time required for applying data assimilation techniques to these high dimension models. The second solution is therefore applied for nowcasting up to two hours. Such nowcasting system involves two main parts. The first one estimates the motion field and the second one forecasts the rainfall location and quantity using the radar images and the estimated motion field. The paper focuses on the estimation process, which is obtained by assimilating radar images with an image model describing the temporal evolution of the studied system. The state vector of this model includes the motion field, the image brightness and a description of the objects displayed on the image data. These objects correspond to the rainy regions, whose displacement must be correctly assessed if the estimated motion field is accurate. Results demonstrate that this extended description of the image data allows to produce an improved forecast.