An occlusion tolerent method for multi-object tracking

A novel approach is proposed for tracking occluded vehicles in real road traffic environment. The third-level DWT is inducted into background difference to filter the high frequency noises in the background, decrease the resolution of frame and detect motion regions. A scene model and a box adjustment model are built to help predict and interpret the occlusion and tracking procedure. A 2-dimensional bounding box is used to represent vehicle, Kalman filter and weighted scale parameter to predict whose center position and motion. The final predictive vehicle-region is estimated using partial observations and adjusting unobservable bounding edges. The center of this new vehicle-region will be feedback to the Kalman filter as the measurement to modify the Kalman gain and estimate the optimum position. The proposed method is an efficient tracking algorithm, as shown in some experiments with real traffic scenes.