Optimization of data access in GPGPU based implementations of Spatio-Temporal Track-Before-Detect algorithm

Tracking systems need accurate and fast computations. Track-Before-Detect (TBD) algorithms are applied if the signal is weak (e.g. SNR<1). Spatio-Temporal TBS algorithm uses IIR-based architecture and there are two filtering parts related to the multidimensional processing. Computation of TBD algorithms needs for example GPGPU devices. The main limitations of GPGPU are memory bus bottleneck and limited availability of fast internal shared memory. A few memory optimization techniques are presented. The new technique based on the dual representation of the state space data in proposed. The compression technique allows application of 8-bit representation for state space and floating point state space together. Floating point values are used for areas where the object is located. The TBD system tracks objects and predicts signal data representation together.