Design And Implementation Of Eye Tracking System Based On ARM

As an important research part of computer vision,eye-tracking technology has broad application in the fields of human-computer interaction,smart healthcare,and advertising.However,the traditional eye-tracking system usually relies on professional hardware devices such as eye tracker,which is expensive to use and hinder popularization of the technology.This thesis proposes a desktop eye-tracking system solution by image processing,which studies the implementation on ARM(Advanced RISC Machine)architecture by formulate software procedures and design image processing algorithms.The system is based on the Raspberry Pi 4B embedded computing platform.It can obtain real-time images from USB camera,and then extract face from the real-time image,and finally locate and track the eye center continuously.The hardware of the system is low-cost,and it has certain generality and portability.The main research work in the thesis includes the following aspects:First,the principle of the system is analyzed,and the overall system design is divided into four parts: the design of face detection algorithm,the design of eye state recognition algorithm,the design of eye-tracking algorithm,and the implementation and test of the overall system.In order to achieve better performance,Open CV(Open Source Computer Vision Library)and Tengine(inference engine for embedded device)are used as image processing tools,and the Qt framework is used to develop the GUI(Graphical User Interface)of the system.Then the software part of the system is designed,which is divided into two parts: the computer side and the ARM side.The main idea is to implement and test the algorithm on the computer side with higher performance,and then ported the algorithm,designed the GUI and implemented the overall system on the ARM side.Then,the algorithms of the system are designed.First,design the face detection algorithm.After analyzed the principles of the Haar-Ada Boost(Haar-like Adaptive Boosting)algorithm and the Retina Face algorithm,tested the performance of the two algorithms on the Caltech 10,000 Web Faces dataset and real-time image,the results show that the accuracy rate of the Retina Face is higher than the Haar-Ada Boost,and the time per frame of the Retina Face is also lower than the Haar-Ada Boost,then the Retina Face is selected as the face detection algorithm of the system.Then design the eye state recognition algorithm.Inspired by the circular feature of the eye,fast radial symmetry transformation is used to perform calculations on the eye area image,and recognized the user’s eye state according by the transformation result shows unimodality when the eye is opened.The test of real-time image show that the accuracy of the algorithm has reached 93.758%.Then design the eyetracking algorithm.The algorithm traverses each point in the image,calculates the displacement vector between that point and other points,and then calculates the dot product of the displacement vector and the gradient vector of other points,then finds the eye center position with the most vector intersections.Considered the performance of the hardware platform,designed a multithreading dot product acceleration calculation method.The test of the Bio ID dataset show that the algorithm has83.6% accuracy,and the test of real-time image show that the algorithm has 28.15 FPS.Finally,completed the build environment,then implemented and tested the overall system on the ARM side.The implementation and test results of the system on different platforms show that the eye-tracking system based on the ARM platform has obvious advantages of low power consumption with low-cost,and the overall frame rate when in use has reached 10.31 FPS,which proved that this system does have certain practical value.