Forming Adaptive Machine Learning Object Position Measurement System in Monocular Vision Environments

• Main Authors: WU, CHENG-XUAN, 吳丞軒
• Other Authors: Horng, Gwo-Jiun
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/wyyxcs

碩士 === 南臺科技大學 === 資訊工程系 === 107 === Forming Adaptive Machine Learning Object Position Measurement System in Monocular Vision Environments is a robot system that implements object detection and an algorithm to calculate objects’ distance. By defining parameters such as camera height, tilt angle and focal length, as well as initializing the parameters of the environment, the robot was able to calculate the correct objects’ distance from an input image. A caveat of this, however, is physical variability in the environment space. Any modification of the parameters set prior, (such as camera tilt, angle, etc.) will result in a fallacious calculation of an objects’ distance. To correct this, this paper proposes an adaptive correction method which combined an accelerometer and using machine learning regression to learn the function between robot tilt angle and output of accelerometer in order to derive the modified parameters. After training the regression model then using it to updates the angle parameter in runtime. The results revealed that the system allowing the robot to provide a stable and correct calculation of objects’ distance even the robot's posture has changed.