Real-Time Reconfigurable Processor to Detect Similarities in Compressed Video Using Generalized Hough Transformation

In this work, we present a FPGA-based Generalized Hough Transform custom processor to calculate similarities between arbitrary shapes. Raw data are 44 x 36 DC images extracted directly from low-resolution compressed video (352 x 288). The outputs are two numbers per frame that quantify the image similitude in terms of scale and rotation. The proposed architecture efficiently resolves the detection of pixel pairs, and the voting of distances and rotations, without memory access conflicts. These operations are inherent to Hough transformation. The paper condenses some circuit solutions suitable to hardwiring video processing. They take full advantage of using small embedded memories as look-up tables. The complete processor is validated with benchmark video samples that cover different scenarios and problems: sport, drama, and news. The final version internally operates at 100 MHz and fits inside a small FPGA chip. The highly concurrent architecture employs both pipelining and parallelism using hardware replication. The final performance is over 40 Giga fixed-point operations per second.