A Comparison of Architectures for Track Fusion

Fusion on track level is desirable for automotive perceptions systems since it enables the use of distributed architectures and communication networks with limited bandwidth. In contrast to fusion of memoryless inputs, input tracks are subject to various correlations, violating the uncorrelated input assumption of standard linear estimators. This paper presents an overview of approaches to deal with correlated inputs together with an evaluation of the performance, with respect to state estimation accuracy, of three standard fusion methods for track fusion. The selection is restricted to methods with low computational demands and compatible with inputs from vehicle integrated sensor modules. All three methods perform the tested tasks well. The results show the advantages and drawbacks of the three architectures for various scenarios and that a final architectural decision should account for more parameters.