**6. Conclusions**

This paper aimed to design and implement a motorized wheelchair that can be controlled via eye movements. This is considered as a pivotal solution for people with complete paralysis of four limbs. The starting point for this project was a joystick-controlled wheelchair; all other building blocks were designed from scratch. Hardware-wise, the design stage aimed to build an image acquisition system and to modify the existing controller to move the wheelchair based on commands given from users' eyes instead of the manual joystick. The core and most challenging part of this prototype was to build a gaze estimation algorithm. Two approaches were followed: machine learning-based algorithms and template matching-based ones. Both approaches showed extremely high accuracy for a certain lighting condition. Nevertheless, template matching did not have a robust performance, as the accuracy had dropped dramatically when it was tested in another lighting condition. The Convolutional Neural Networks had superior performance in this regard, and were therefore the chosen technique for gaze estimation. Based on Table 3, the proposed system is very suitable for real time application. The classification accuracy of the gaze estimation is 99.3% which is above 95%, i.e., the norm for such type of applications. The model can process 99 frames per second and can make a prediction in ~1.57 ms, which 5x faster than the real-time requirements and the recent works reported in the literature. The safety of the user was given a paramount concern, and was ensured by incorporating an array of five ultrasonic sensors to cover the whole range of motion of the wheelchair. It can instantly stop if any obstacle is detected in this range. The system can be further developed by opting for a global approach rather than a user specific one, to eliminate the need for a calibration phase. Moreover, the current gaze estimation algorithm can stop the wheelchair or make it move only in three directions, namely, forward, right, and left. However, if a regression was used instead of classification, more degrees of freedom could be provided for movement. The authors are investigating on the joystick controller protocol to replicate it in the microcontroller used for joystick control. In this way, a non-mechanical system of a joystick controller can be developed to control the wheelchair controller directly.

**Author Contributions:** Experiments were designed by M.E.H.C., A.K. and S.K. Experiments were performed by M.D., T.R., K.A.-J. and A.H. Results were analyzed by M.D., M.E.H.C. and S.K. All authors were involved in the interpretation of data and paper writing. All authors have read and agreed to the published version of the manuscript.

**Funding:** The publication of this article was funded by the Qatar National Library and Qatar National Research Foundation (QNRF), grant numbers NPRP12S-0227-190164 and UREP22-043-2-015.

**Conflicts of Interest:** The authors declare no conflicts of interest.
