Research On Embedded Floating Objects Recognition System

With the rapid development of our society,a large number of production and domestic garbage is produced every day.Much garbage is thrown into waters,such as rivers,lakes,seas and etc.The water pollution is becoming more and more serious,especially the impact of floating garbage.At present,the way to clean up the floating garbage mainly depends on the manual or artificial auxiliary operations,which cannot realize automatic identification and cleaning of floating garbage,so that cleaning floating garbage is inefficiency and the cost is also very high.The main limitation is that the water environment is complex,and automatic identification and tracking the floating garbage is difficult.Recently,machine learning obtained fast developing,so it can be used to solve the problem of automatic identifying and tracking the floating garbage.According to the above motivation and the actual requirements,this thesis designs an embedded recognition system in which the floating object identification algorithm can be migrated and the real-time performance can be ensured.The research works in this thesis mainly contains:(1)The image style transfer algorithm is applied to identify the floating objects.The pre-trained network model is retrained,and the algorithm can learn the water style in the image,which can be used for comparison of the style characteristics of the detected image.In order to optimize the algorithm,the appropriate parameters are designed so that the real-time recognition performance is improved.By using OpenCV to the gray pattern image which is obtained from the image style transfer algorithm,the disturbance in the image can be eliminated and the floating object can be recognized.(2)The binocular stereo matching algorithm is used to achieve the target tracking by using the binocular cameras which are calibrated through Zhengyou Zhang calibration method.Moreover,the recognition results are used as the tracking template of KCF algorithm to realize the tracking function of floating objects.(3)The embedded floating object recognition system was built mainly by the hardware platform of Raspberry Pie 4B and two Logitech c920e cameras.Opencv3.4.3 and Tensorflow learning framework are transplanted to the embedded Linux operating system Raspbian Buster to realize the real-time recognition and tracking algorithm of floating objects.