Detection of pedestrians aboard a mobile vehicle by machine learning
(…)
When it was contacted in late 2018 by a French metallurgist who wanted to install pedestrian detection cameras on its mobile machines, Yumain, a French start-up, chose machine learning algorithms based on spike, an artificial neural network that requires less analysis than deep learning. “A vision system based on deep learning performs a complete analysis of the image: if you have a million pixels, it will process each of them,” says Marc Benoit, director of Yumain. “With spike, the algorithm only analyzes the elements that evolve, the movement in the image“. After six months of testing on a pilot site, then three, the metallurgist is “in the process of deploying” Yumain’s box in all its factories, adds the CEO.
(…)
If the detection boxes seem to convince their customers, the two start-ups have not given up on integrating embedded deep learning in the near future. Marc Benoit is banking heavily on advances in microelectronics.
“Yumain’s teams are developing, with the CEA, their own Asic, a hyper-specialized chip on which we will be able to embed our own neural networks“. (…) “The Holy Grail,” Marc Benoit dreams, “would be an ultra-low power Asic, which would consume less than 1 nanowatt, compared to 3 to 5 watts for the Asic we are working on and 15 watts for our current GPUs“.