Abstract: Provided is a neural architecture search method based on knowledge distillation, which trains a student network using knowledge acquired from a teacher network and searches a target neural network. The neural architecture search method may include the steps of: (a) extracting an image feature map from a learning model of the student network; (b) calculating a loss function by comparing an image feature map extracted from a learning model of the teacher network to the image feature map extracted in the step (a); (c) selecting a block whose capacity is to be increased and a block whose capacity is to be decreased, for each learning model block of the student network, based on the loss function; and (d) deciding a candidate learning model of the student network according to the block architecture selected in the step (c).
Abstract: Disclosed is a lane departure warning determination method using driver state monitoring, which includes (a) recognizing a lane around a vehicle from an image photographed by a lane recognition camera and extracting lane information; (b) extracting facial feature points of a driver from an image photographed by a driver recognition camera; (c) determining from the lane information whether the vehicle is in a lane departure state or a lane approaching state; (d) when it is determined in step (c) that the vehicle is in the lane departure state, generating a warning when a face tilt of the driver exceeds a first angle; and (e) when it is determined in step (c) that the vehicle is in the lane approaching state, generating a warning when the face tilt of the driver exceeds a second angle.
Type:
Application
Filed:
June 18, 2020
Publication date:
January 7, 2021
Applicant:
A.I.MATICS Inc.
Inventors:
Sang Mook LIM, KWANG IL PARK, Jin Hyuck KIM