Gaussian mixture algorithm （GMA） is an effective approach for off-road terrain estimation, but still suffers from some difficulties in practical applications, such as complex calculation and object abstraction. In this paper, GMA is modified to improve its real-time performance and to provide it with a potential ability of obstacle detection. First, a selection window is designed based on the dominant-ellipse-principle to limit the probability distribution area of each measurement point, therefore avoiding the calculation on the cells outside the dominant ellipse. Second, a clustering approach is proposed in order to distinguish objects efficiently and decrease the operation area of one laser scan. Third, a virtual point vector is introduced to further reduce the computational load of the mean square error matrix. The modified GMA is experimented on a tracked mobile robot, and its improved performance is shown in comparison to the original GMA.
