@inproceedings{oai:ipsj.ixsq.nii.ac.jp:00210332,
 author = {Xin, Du and Yutaka, Shinkai and Mizuki, Itoh and Yoshiki, Yamaguchi and Xin, Du and Yutaka, Shinkai and Mizuki, Itoh and Yoshiki, Yamaguchi},
 book = {Proceedings of Asia Pacific Conference on Robot IoT System Development and Platform},
 month = {Mar},
 note = {This study proposes an approach to estimate human action in real-time by analyzing sensing data from multiple accelerometers with FPGA. The body action distinguished by sensing data is estimated by a neural network called a self-organizing map (SOM). It produces a low-dimensional representation of data sensed by multiple accelerometers using unsupervised learning. It can visualize them in the learning space automatically and classify humans' actions with high accuracy. However, the iterated computation on SOM requires many computational efforts, which requires choosing efficient computational chips. In this study, FPGA was chosen to develop a computational technique because the spatial parallelism on FPGAs is useful to implement SOM parallelism. In our experiments, the trial system comprises one Xilinx Spartan-6 FPGA, a small FPGA, and five multiple 9-axis sensors. Although it was small and straightforward, it could distinguish five actions: walk, run, stand, stair-up, and stair-down., This study proposes an approach to estimate human action in real-time by analyzing sensing data from multiple accelerometers with FPGA. The body action distinguished by sensing data is estimated by a neural network called a self-organizing map (SOM). It produces a low-dimensional representation of data sensed by multiple accelerometers using unsupervised learning. It can visualize them in the learning space automatically and classify humans' actions with high accuracy. However, the iterated computation on SOM requires many computational efforts, which requires choosing efficient computational chips. In this study, FPGA was chosen to develop a computational technique because the spatial parallelism on FPGAs is useful to implement SOM parallelism. In our experiments, the trial system comprises one Xilinx Spartan-6 FPGA, a small FPGA, and five multiple 9-axis sensors. Although it was small and straightforward, it could distinguish five actions: walk, run, stand, stair-up, and stair-down.},
 pages = {71--72},
 publisher = {情報処理学会},
 title = {A study of FPGA-based real-time action estimation with multiple accelerometers},
 volume = {2020},
 year = {2021}
}