PercevalCam a Smart Camera for Computer Vision and Deep Learning for Robotics

AI and more particularly the Deep Learning is becoming pervasive when perception is involved. Indeed, in the recent years, DNNs have succeeded to enhance the state-of-the-art with a quality that was not even expected. This revolution can be partially granted to the introduction of large GPUs. This technology allowed to compute, very quickly and in parallel, the training phase and the inference phase of deep learning algorithms. It can be stated that GPU technology is the foundation that allowed the maturity of DNN. However, in situations where these large GPUs are not available, which is mainly occurring on the edge, it is currently not possible to obtain real-time results that are not a trade-off relatively to the quality. This point is even more sensitive for the industrial sector where real-time means having a solution in only few milliseconds. Indeed, many industrial systems like robots have to answer reliably in a very short period of time. Usually, these systems cannot embed voluminous computer, can be limited in term of battery life, or simply want to limit the global cost.