Abstract: Based on the Kalman predictive model and orthogonal projection locating theory, a novel algorithm is proposed to segment the object from overlapped vehicle images, which is partly hidden by other object(s). In the algorithm, a segmentation window is framed for a specific occluding vehicle in binary image by means of the vehicle's historically motion information in time domain and Kalman predictive result, and then the minimal circumscribed rectangle of the vehicle is determined using horizontal-vertical orthogonal projection and dynamic threshold theory. A function is designed to evaluate if the segmentation is reasonable with corresponding results given. Experimental results showed that this algorithm can effectively segment overlapped vehicles from each other and track them steadily with low computational complexity available to meet the requirements for real-time system.