Measurement fusion for target tracking under bandwidth constraints

In multi-sensor estimation/tracking-such as for radar or sonar surveillance systems-direct sharing of measurements is preferable to that by sharing of local estimates, since in the former case some optimal sort of fusion is possible. A recent study has shown that such a scheme is competitive in terms of bandwidth. Since this study was quite conservative in its assumptions about quantization, we explore the issue further, to see if such a high bandwidth (number of bits per measurement) is really necessary and whether improvement is achievable via intelligent quantization. We approach the problem both practically and theoretically. We discover that there is substantial improvement possible: indeed, as opposed to the 8, 16, or even 32 bits per dimension of current expectation, it is reasonable that only 2-3 bits per measurement, per dimension, be used. Further, through use of a quantization scheme in which resolution is fine where a target is expected to be and coarse elsewhere, quite significant improvements are available. Such a quantizer is simple to implement via companding, and both error-function and μ-law (logarithmic) companding appear to work quite well. In cases that the measurement is accurate compared to the prior information-that is, for a highly-maneuvering target-the performance gain can be dramatic