Real-time side vehicle tracking using parts-based boosting

This paper presents a real-time vision-based side vehicle detection system employing a parts-based boosting algorithm. Working at the part level, as opposed to the whole object level, enables a more flexible class representation and allows scenes in which the query object is significantly occluded to be classified. Therefore, a parts-based learning approach is proposed in order to better deal with side vehicle variability, illumination conditions, partial occlusions, and rotations. Most existing boosting learning algorithms usually select weak classifiers by minimizing a cost directly associated with the error rate, where the learned strong classifier may be sub-optimal for applications in terms of error rate. Nevertheless, the proposed Adaboost approach selects weak classifiers by minimizing multiple types of error functions. The idea is to define multiple types of error functions based on current strong classifier and each selected weak classifier results to represent different effects of each weak classifier. Therefore, weak classifiers can be selected with different requirements at the same time to avoid a sub-optimal solution. To reduce system computation, window-based tracking is employed. Moreover, Kalman filtering is used to predict the position of each part of vehicles in the image plane to effectively relocate the tracking windows. Compared with existing approaches, the proposed system appears to be capable of improving system efficiency and accuracy under varying lighting conditions, changing vehicle poses, and in the presence of partial occlusions. Our approach has been successfully validated in real traffic environments by performing experiments with a CCD camera mounted onboard a highway vehicle.