A research group led by Professor Yasushi Yagi of the Institute of Scientific and Industrial Research, Osaka University has developed a high-precision gait authentication technology using deep learning, which is attracting particular attention among AI (artificial intelligence) technologies.
People's walking style (gait characteristics) does not change depending on their clothes and hairstyle, and can be extracted from low-resolution images taken from a distance with a security camera or the like.It is a very practical feature for personal authentication.However, if the walking direction of the person is different with respect to the camera, the way the person sees is greatly different, so it is difficult to authenticate the gait with the conventional technology.Therefore, it is effective to compare the differences in abstract features such as the way the hands are waved and the width of the legs.
This time, we proposed a unique deep learning model, and by properly using these features, we made it possible to perform highly accurate gait authentication from images of people walking in different directions.As a result, when the walking direction is significantly different, the error rate of personal authentication was about 40% in the conventional technology, but it was reduced to about 4%, which is the highest accuracy in the world, with the technology developed this time.This will greatly expand the scope of application of the "Future Science Investigation" gait appraisal.
Furthermore, by appropriately changing the evaluation criteria used for deep learning, it is possible not only to authenticate the person but also to identify a specific person from a plurality of people captured by the camera.By using these technologies to analyze the movement route of the same person in stores and commercial facilities, it is expected to be applied to various uses other than criminal investigation, including marketing applications such as service provision according to customers. NS.
In Japan, the modernization of safety measures against terrorism and crime is an urgent issue ahead of the Tokyo Olympics in 2020, and by using this technology, it is possible to quickly identify and track suspicious persons and suspects using security cameras. Expected to do.
Paper information:[IEEE Transactions on Circuits and Systems for Video Technology] On Input / Output Architectures for Convolutional Neural NetworkBased Cross-View Gait Recognition
