This paper provides a sensor fusion approach to improve the accuracy in building an occupancy map based on an established ultrasonic sensor model. We first construct a sensor model for the ultrasonic range finders mounted on the Pioneer P3-DX mobile robot by experimental test and mathematical reasoning. Based on the sensor model established, a probability can be assigned to each cell of the grid map that represents the possibility of obstacles. Second, by using sensor fusion technique, like Dempster-Shafer theory, probabilities of the cells can be combined to derive a new probability for representing the cell. As mobile robot moves, the probability of cell will be updated continuously. Finally, a complete occupancy grid map for the unknown environment can be established for use in navigation of the autonomous robot.