US 10,890,916 B2
Location-specific algorithm selection for optimized autonomous driving
Kye-Hyeon Kim, Seoul (KR); Yongjoong Kim, Pohang-si (KR); Hak-Kyoung Kim, Pohang-si (KR); Woonhyun Nam, Pohang-si (KR); SukHoon Boo, Anyang-si (KR); Myungchul Sung, Pohang-si (KR); Dongsoo Shin, Suwon-si (KR); Donghun Yeo, Pohang-si (KR); Wooju Ryu, Pohang-si (KR); Myeong-Chun Lee, Pohang-si (KR); Hyungsoo Lee, Seoul (KR); Taewoong Jang, Seoul (KR); Kyungjoong Jeong, Pohang-si (KR); Hongmo Je, Pohang-si (KR); and Hojin Cho, Pohang-si (KR)
Assigned to STRADVISION, INC., Pohang-si (KR)
Filed by STRADVISION, INC., Pohang-si (KR)
Filed on Dec. 31, 2019, as Appl. No. 16/731,083.
Claims priority of provisional application 62/798,821, filed on Jan. 30, 2019.
Prior Publication US 2020/0241544 A1, Jul. 30, 2020
Int. Cl. G05D 1/00 (2006.01); G05D 1/02 (2020.01); B60W 50/08 (2020.01); G06N 3/08 (2006.01)
CPC G05D 1/0221 (2013.01) [B60W 50/085 (2013.01); G05D 1/0219 (2013.01); G05D 1/0278 (2013.01); G06N 3/08 (2013.01)] 30 Claims
OG exemplary drawing
 
1. A learning method for performing a seamless parameter switch by using a location-specific algorithm selection, to thereby allow a subject vehicle to perform an optimized autonomous driving in each of regions, comprising steps of:
(a) instructing, by a learning device, when one or more K-th training images corresponding to at least one K-th region are acquired, at least one K-th convolutional layer of a K-th Convolutional Neural Network (CNN) included therein to apply at least one K-th convolution operation to the K-th training images, to thereby generate one or more K-th feature maps;
(b) instructing, by the learning device, at least one K-th output layer of the K-th CNN to apply at least one K-th output operation to the K-th feature maps, to thereby generate one or more pieces of K-th estimated autonomous driving source information;
(c) instructing, by the learning device, at least one K-th loss layer of the K-th CNN to generate at least one K-th loss by using the K-th estimated autonomous driving source information and its corresponding K-th Ground-Truth (GT) autonomous driving source information, and then to perform backpropagation by using the K-th loss, to thereby learn at least part of K-th parameters of the K-th CNN;
(d) storing, by the learning device, the K-th CNN in a database after tagging K-th location information of the K-th region to the K-th CNN,
wherein the K-th CNN is one of a first CNN to an M-th CNN, and wherein M is an integer larger than 1, and K is an integer from 1 to M;
(e) by the learning device, (i) acquiring information on each of a (K_1)-st CNN to a (K_N)-th CNN, selected among the first CNN to the M-th CNN, corresponding to each of a (K_1)-st region to a (K_N)-th region whose distance from the K-th region is smaller than a first threshold, and then (ii) calculating similarity scores including a (K_1)-st similarity score between the K-th parameters and (K_1)-st parameters of the (K_1)-st CNN to a (K_N)-th similarity score between the K-th parameters and (K_N)-th parameters of the (K_N)-th CNN; and
(f) generating, by the learning device, generating a K-th representative CNN corresponding to a K-th extensive region including the K-th region by referring to at least part of (i) one or more specific CNNs, among the (K_1)-st CNN to the (K_N)-th CNN, whose similarity scores are larger than a second threshold and (ii) the K-th CNN, and
wherein N is an integer larger than 0.