Scalable object detection accelerators on FPGAs using custom design space exploration

We discuss FPGA implementations of object (such as face) detectors in video streams using the accurate Haar-feature based algorithm. Rather than creating one implementation for one FPGA, we develop a method to generate a series of implementations that have different size and performance to target different FPGA devices. The automatic generation was enabled by custom design space exploration on a particular design problem relating to the communication architecture used to support different numbers of image classifiers. The exploration algorithm uses content information in each feature set to optimize and generate a scalable communication architecture. We generated fully-working implementations for Xilinx Virtex5 LX50T, LX110T, and LX155T FPGA devices, using various amounts of available device capacity, leading to speedups ranging from 0.6x to 25x compared to a 3.0 GHz Pentium 4 desktop machine. Automated generators that include custom design space exploration may become more necessary when creating hardware accelerators intended for use across a wide range of existing and future FPGA devices.