應用影像深度學習於人形機器人足球賽之即時物件偵測
Other Title
Utilizing Visual Deep Learning for Large-range Object Detections of Humanoid Robot Soccer Games
Journal
科儀新知
Journal Issue
231
Start Page
11
End Page
26
ISSN
1019-5440
Date Issued
2022-06
Author(s)
林昆鋒
Abstract
RoboCup為全球人形機器人之重要比賽之一，其中雙足人形機器人足球賽除了需考慮到雙足動態步行之穩定性外，也需進行即時物件偵測、辨識與定位，以作為自主決策與運作之依據。所探討之人形機器人搭載NVIDIA Jetson TX2嵌入式人工智慧運算器實現機器人視覺即時偵測、辨識之運算，並以深度學習卷積神經網路開發，導入You Only Look Once網路進行足球與機器人類別之偵測與辨識。此一成果在RoboCup 2017賽事進行實際驗證，並獲得中型組人形機器人足球賽第二名成績。
RoboCup is one of the most important humanoid robotic competitions in the world. The development of humanoid soccer robot considers not only the bipedal locomotion stability, but also the real-time object detection, identification and positioning which help the autonomous decision-making and operation during competition. The proposed humanoid robot is equipped with a NVIDIA Jetson TX2 embedded artificial intelligence controller to realize the aforementioned real-time object detection based on the deep convolutional neural network (CNN). Practically, the classes of soccer ball and humanoid robots are detected and recognizes in terms of the You Only Look Once (YOLO) model. The results are validated in the RoboCup 2017, and our team won the 2nd place of the TeenSize league.
Type
journal article