Qunshan He (何群山)

    In the RoboMaster Robotics Competition, participants need to control their robots to attack enemy’s robots by launching projectiles. And the only effective attacks are made by striking the Armor module, which is highly characterized by two bright light-bars (red or blue). We come up with a machine vision method to recognize enemy’s armor and aim at the targeted one to realize robot automatic attack.
    First we extract specific color area to construct light-bar instance, then we use spatial feature to construct armor instance, next we crop the armor area as the input of a SVM model to classify the corresponding robot. Finally we calculate the yaw and pitch angles based on perspective transformation principle to control the robot aiming at the targeted robot.
    This algorithm has been running in our robots and shows nice performance with high accuracy (95%) and speed (200~300fps). And it has also been open sourced in GitHub and Gitee.

    Standard microscope is limited by the Numerical Aperture (NA) of physical lenses, if we want to see a feature with a higher resolution, we must switch to a different microscope objective, with a reduced field of view.
    However, by applying LED matrix as Illumination, combined with computational optics, we can boost the resolving power of standard microscopes by more than an order of magnitude.
    Here I develop a computational optical microscope based on the concepts of Aperture Synthesis and Phase Retrieval, realizing wide-field and high-resolution quantitative phase imaging, for which I also proposed a Dynamic Sampling Denoise Algorithm based on LMS.
    This microscope is currently applied to optical lens inspection to present tiny and transparent defects of optical lens.