FPGA implementation of mono and stereo inverse perspective mapping for obstacle detection

This paper presents the implementation of real time Inverse Perspective Mapping algorithms on a FPGA. The IPM is a spatiotemporal algorithm that allows the detection of obstacles from two images I₁ and I₂, either acquired by a moving camera, or by a stereoscopic sensor, knowing the relative positions between the two view points. In the case of mono Inverse Perspective Mapping (MIPM), the homography is based on movement information (odometry) while in the case of stereo-inverse perspective mapping (SIPM), the homography depends on the inter-ocular distance. For MIPM, a co-design approach is proposed to compute the homography matrix using a soft-processor. For SIPM, a pre-computed table that encodes simultaneously the homography transformation and the lens distortion correction is used. We present an optimized architecture to implement a real time homographic transformation using a SDRAM memory, which is cheaper than the SRAM. The algorithms are intended to work at least at 30 fps with a VGA image resolution, they are evaluated using two different platforms: a Xilinx Virtex 6 development kit for MIPM, and Altera Cyclone 2 based custom board for SIPM.