Abstract:
Localization and mapping are essential tasks for autonomous vehicles. Several research projects focused on the design of embedded systems executing simultaneous localization and mapping algorithms (SLAM). In this work, we propose a generic prototyping platform dedicated to SLAM, based on a "Meet-in-themiddle" approach. Two key paradigms have been used for this: SoCs and open-source. The platform has two parts: hardware part and software part. The hardware part is designed around a low cost SoC-FPGA integrating a CPU and an FPGA in the same chip. The accelerator circuit on the FPGA is designed using a high level language: OpenCL. The use of this language reduces development time and offers the possibility of porting to other hardware acceleration platforms. The Software part is based on the Linux system and open-source bricks. A build automation tool is used to build and generate the embedded Linux system. Such a tool allows a reduction in development time as well as retargeting using cross compilers. At the end, a case study on a visual SLAM was carried out. An existing implementation of SLAM has been implemented on our platform. A software / hardware co-design methodology is used. This methodology makes it possible to decide which SLAM function will be executed on which part (software or hardware) of the platform, based on profiling and analysis of the SLAM code. OpenCL kernel optimization techniques were used to increase performance. The operation of the SLAM was in real time (57.14Hz), with an acceleration of 1.93 times the software implementation.