SW/HW implementation of image covariance descriptor for person detection systems

Due to the complexity and the high performance requirement of person detection application, the design of embedded systems is the subject of different types of design constraints such as execution time, time to market, energy consumption, etc. Some methodologies were proposed in order to satisfy the different design constraints. This paper presents a mixed HW / SW design methodology for the implementation of an image covariance descriptor on a reconfigurable Altera StrtixIII FPGA platform. This implementation is intended for person detection and re-identification systems. It begins with a pure software implementation of the descriptor. A dynamic profiling is then made to identify the critical functions that require a hardware implementation. Finally, the most time consuming functions are implemented as hardware accelerators.