Extended Object Tracking via Positive and Negative Information Fusion

In this paper, a new information fusion and state estimation algorithm is proposed for extended object tracking. An extended object has a spatial extent that is relatively larger than a sensor's resolution, and thus, multiple measurements are obtained from the object at each sensor scan. For example, in lidar-based tracking applications, a car can be considered as an extended object where a lidar sensor measures multiple laser lights reflected from various points on the car's surface. The extended object tracking is a challenging task since multiple measurements from the object need to be systematically assimilated to estimate the shape as well as the kinematic states (e.g., position and velocity) of the object. In particular, the multiple measurements along with sensor-to-object geometry provide positive (where the object exists) and negative (where the object should not exist) information, and these two types of information need to be incorporated in the estimation process. In the proposed algorithm, novel measurement models are developed for each type of information based on Gaussian process regression and virtual measurement generation. Based on those models, the extended object tracking problem is formulated as a Bayesian inference problem for which the unscented Kalman filter is used for its robustness and efficiency. The effectiveness of the proposed algorithm is demonstrated with illustrative car tracking examples.