FPGA implementation of event-based optical flow for robotic applications

The speed with which technological evolution has reached increasingly extraordinary goals, in recent years, has now opened the doors to the possibility of develop more precise and refined devices with respect to the recent past. In this way, the research in the artificial intelligences field represents a fundamental joint for both possible future implications in the industrial development and in the health world. The purpose of this work is to build up a reconfigurable hardware solution, able to interface with modern vision sensors emulating the behaviour of the human retina. This final product, at the end of a close cooperation with the Istituto Italiano di Tecnologia (iit), is significantly useful to have a base-test for the optical flow computation. This could represent, furthermore, an important source of informations from an AI perspective or for feasible biological involvements. The path taken through this elaborate starts from a close analysis on neural networks, learning systems and the different vision sensors currently in commerce. The frame-based sensors are the oldest solution available, but present critical problems in terms of redundancy, latency, long time intervals and so on. For that reason all the efforts in this direction are focused on the event-based solution. These last have, indeed, excellent operating characteristics that make them suitable for human-like interfaces. Generally, the movement of an object generates a variation of brightness in the reference field of a sensor: through a sleek algorithm, which is the case-study of this thesis, it is possible to obtain spatio-temporal informations about the event related to the movement itself. In this way, precise visual flow orientation and amplitude can be estimated. The low observation interval also makes event-based sensors the best for this purpose compared to frame-based one. Starting from the analysis of the reference paper, the operating algorithm was extracted and reproduced in floating-point version in a parametrized and reconfigurable Matlab script. Subsequently, this function was also adapted in fixed-point, offering solutions to some interface and overflow problems coming out from the sensor. By comparing the results obtained by using a real dataset, kindly provided by the iit, it was possible to make a statistic estimation of the errors in the modelling phase. This represents the starting point for the subsequent RTL design phase of the device. All the solutions adopted are described in detail. The reference SoC module is a Trenz Electronic TE0715-04-30-1C, which integrates a Xilinx Zynq XC7Z030-1SBG485C. At the end, a comparison between the expected results in Matlab and the ones obtained from the hardware simulation, as well as the informations coming from the synthesis of the device in terms of area, power consumption and timing have been realized for different possible configurations.