A Low-Cost Passive Stereo Vision Solution for an Autonomous Wheelchair

The field of ADAS (Advanced Driver-Assistance Systems) is nowadays seeing an increase in interest from the scientific community, due to its numerous applications and possibilities in many engineering contexts, such as automotive, robotics, and so on. Considering this, the topic of computer vision applied to autonomous driving is therefore a fundamental branch of the research. The main purpose of this work is to develop and validate a stereo vision system that is suitable for running in an efficient manner on an embedded system (Nvidia Jetson Nano) with real-time constraints. The context is the autonomous navigation of a mobile robot in the healthcare segment, and in particular of an autonomous driving wheelchair developed by Teoresi Group and known as ALBA project. The baseline for the work is the active stereo camera system developed by Intel (model D435i) and currently employed by the robot, that embeds an infrared (IR) emitter along with the stereo cameras and is therefore efficiently handling most of the various conditions that may come up in the navigation scenario. Along with the performance requirements, the developed algorithm must be consistent with the economic needs imposed by the commerciality requirement of the product, which the active stereo exceeds. In order to reach the abovementioned objectives, the methodology has included firstly a study of the state-of-the-art and a search for suitable approaches and algorithms to the stereo matching problem, and secondly a phase of testing and of integration in a software pipeline able to deliver a data structure usable for navigation purposes, i.e., a cost map. The most suitable algorithm resulted to be a version of the Semi-Global Matching (SGM) approach, optimized for usage on embedded GPUs. The validation phase, performed in ROS2 (Robot Operating System 2) environment, has been completed by testing the algorithm in various environmental conditions and comparing its performances with those of the active stereo, subsequently highlighting the strengths of a passive stereo and the scenarios in which additional sensors are instead required. The experimental analysis has shown that the developed equipment is much more convenient than the active stereo apparatus from the economic viewpoint, while its performances are sufficiently near to the objective ones, depending on the real world scenario. In particular, the passive stereo resulted robust against reflection effects that are ambiguous to IR-based systems, while it achieved bad performance in low lighting conditions or highly untextured areas, as foreseen by the stereo vision theory.