●Technique highlights:
1. CNN model architecture optimization techniques:
Propose an Agilev4 architecture, with a composite performance index of mAP＠50:95 and FPS＠FP16 product up to 253.05
2. GOP-mode (Group of Picture) acceleration scheme for real-time inference:
Employ the GOP-mode and TRACKING algorithms to enhance the processing speed from 2.51 to 30.90 frames per second on Jetson Nano mobile platform.
3. Propose a rapid data labeling system HiTag
●Technical breakthrough:
1. CNN model architecture optimization techniques:
Propose a new model architecture Agilev4 achieving a composite performance index of 253.05 (outperforms 135.24 of Yolov3 , and 163.83 of Yolov4 )
2. GOP-mode acceleration scheme for real-time inference:
Split the video input into I-frames and P-frames. Only I-frames are processed with full object detection, while the P-frames results are obtained by tracking. For implementation of the Agilev4 model on the Jetson Nano platform, the FPS is improved from 2.51 to 30.90, indicating a 1231％ enhancement. AP＠50 performance drops slightly from 85.21 to 84.69.
3. In addition, we have also developed an auto-labeling system to expedite the tedious and time consuming ground truth bounding box labeling.
●Applicable industrial applications：
Mobile service robots, autonomous mobile vehicles, disinfection robots, logistics robots, sweeping robots.
●Specs：1. Frame rate on Jetson Nano reaches 30.9 frames per second for 512*512 input resolution; 2. Object detection accuracy of MS COCO mAP＠50 reaches 58.40％, mAP＠50:95 reaches 35％.
●Market analysis：According to statistics from the International Federation of Robotics (IFR), between 2018 and 2019, global sales of professional service robots reached 11.2 billion U.S. dollars, a growth of 32％, and will continue to grow in the future.

線上展網址：
https://tievirtual.twtm.com.tw/iframe/ffebe167-dde3-4446-8b35-eb414e4cb121?group=23bfb1fa-dd5b-4836-81a1-4a1809b1bae5&lang=en