Exploring FPGA dynamic reconfiguration for sound source localization and tracking algorithms

The ever-increasing complexity of new electronic systems designs, combined with the need for very high performance and low power consumption have driven technology development, methodologies and design flows, including the widespread use of reconfigurable FPGA technologies in systems design. The in-system reconfigurability of FPGAs offers the possibility of reconfiguring the device for different functionality without halting the application. Adding this functionality requires special design flows. This thesis first explores the use of different FPGA reconfigurable design flows, using a sample design, which includes a sensor and wireless transmitter along with the FPGA. Two reconfigurable flows were investigated, Dynamic Partial Reconfiguration (DPR) and Multi- Boot (MB). These flows were also used in FPGA implementations of algorithms for the localization and tracking of sound sources. It is shown that the less-studied Multi-Boot flow has some unexpected advantages over the better-known DPR technique. Applications based on sound source localization and tracking require specific algorithms, which can adapt themselves to disturbing noise or the mobility of the target. This work proposes a novel hybrid algorithm with the ability to adapt its search spectrum according to the target movement. This property allows a good balance between required power processing and real-time response, especially when implemented on a FPGA. Moreover, the algorithm fits quite well on a Dynamic Partial Reconfiguration (DPR) or Multi-Boot (MB) strategy. The approach combines the Generalized Cross Correlation (GCC) and the Delay and Sum Beamforming (DSB) algorithms in such a manner that reductions of more than 80% in DSB computation can be obtained while preserving precision and improving real-time capabilities. The novel algorithm is also applied to a number of application topics such as speaker recognition and multiple sound source tracking that use different localization and beamforming techniques. The performance of the author’s hybrid algorithm in these applications is studied and it is shown to retain its advantages in these cases.