Comparison of tracklet fusion and distributed Kalman filter for track fusion

In track fusion, the measurements of individual sensors for each target are processed to generate local state estimates, which are then fused to obtain the global state estimate for the target. When there is no process noise or the fusion rate equals the sensor observation rate, the standard tracklet fusion or equivalent measurement fusion algorithm computes the optimal centralized estimate by extracting the new information in the local estimates. By using an augmented state (similar in concept to accumulated state density) that includes the states at multiple times, this algorithm also produces the optimal centralized estimate when there is process noise and the fusion rate is lower than the measurement rate. Optimal fusion can also be achieved by the recently developed distributed Kalman filter (DKF) but the local estimates are computed using global sensor models and are not optimal given the local sensor measurements. The covariance debiasing DKF has been proposed to avoid this global dependence. Simulations are used to compare the performance of tracklet and DKF fusion algorithms and their sensitivity to process noise and knowledge of global sensor parameters. The results show that tracklet fusion is close to optimal and DKF with covariance debiasing can handle fairly challenging problems.