Kalman-filter based spatio-temporal disparity integration

Vision-based applications usually have as input a continuous stream of data. Therefore, it is possible to use the information generated in previous frames to improve the analysis of the current one. In the context of video-based driver-assistance systems, objects present in a scene typically perform a smooth motion through the image sequence. By considering a motion model for the ego-vehicle, it is possible to take advantage of previously processed data when analysing the current frame.This paper presents a Kalman filter-based approach that focuses on the reduction of the uncertainty in depth estimation (via stereo-vision algorithms) by using information from the temporal and spatial domains. For each pixel in the current disparity map, we refine the estimated value using the stereo data from a neighbourhood of pixels in previous and current frames. We aim at an improvement of existing methods that use data from the temporal domain by adding extra information from the spatial domain. To show the effectiveness of the proposed method, we analyse the performance on long synthetic sequences using different stereo matching algorithms, and compare the results obtained by the previous and the suggested approach.

► We use data from the spatio-temporal domains to enhance disparity maps.
► Improvements were detected using three different stereo algorithms using four test long image sequences.
► The consistent results encourages keep using data from the spatio-temporal domains.