Á¤º¸°úÇÐȸ³í¹®Áö (Journal of KIISE)
Current Result Document : 1 / 1
ÇѱÛÁ¦¸ñ(Korean Title) |
Feature Pyramid Network ±â¹Ý ¿ø°Å¸® µå·Ð °ËÃâ ¹æ¹ý |
¿µ¹®Á¦¸ñ(English Title) |
Feature Pyramid Network-based Long-Distance Drone Detection Method |
ÀúÀÚ(Author) |
±ÇÁ¤ÀÎ
¼Õ¼ÒÈñ
ÀüÁø¿ì
ÀÌÀÎÀç
Â÷ÁöÈÆ ÃÖÇØö
Jeongin Kwon
Sohee Son
Jinwoo Jeon
Injae Lee
Jihun Cha
Haechul Choi
|
¿ø¹®¼ö·Ïó(Citation) |
VOL 48 NO. 03 PP. 0325 ~ 0333 (2021. 03) |
Çѱ۳»¿ë (Korean Abstract) |
µå·ÐÀÌ º¸ÆíÈµÇ¸é¼ µå·Ð¿¡ ÀÇÇÑ »ç°í¿¡ ´ëÀÀÇϱâ À§ÇÑ °¨½Ã ½Ã½ºÅÛÀÇ Çʿ伺ÀÌ Á¦±âµÇ°í ÀÖ´Ù. µå·ÐÀº ºñÇà ¼Óµµ°¡ ºü¸£¹Ç·Î ¿ø°Å¸®¿¡¼ ¹Ì¸® °ËÃâÀ» ÇØ¾ß ÇÑ´Ù. ÇÏÁö¸¸, ¿ø°Å¸® ¿µ»óÀÇ °æ¿ì ¸ñÇ¥¹°ÀÇ Å©±â°¡ ¸Å¿ì ÀÛ°í º¹ÀâÇÑ ¹è°æÀ» Æ÷ÇÔÇÒ ¼ö ÀÖ¾î ÃÖ±Ù °´Ã¼ ŽÁö ºÐ¾ßÀÇ µö·¯´× ±â¼úÀ» ÀÌ¿ëÇÏ´õ¶óµµ ¿À°ËÃâ·üÀÌ ¸Å¿ì ³ô´Ù. µû¶ó¼ º» ³í¹®¿¡¼´Â ¼ÒÇü ¸ñÇ¥¹°¿¡ ³ôÀº ¼º´ÉÀ» °®´Â ³×Æ®¿öÅ©ÀÎ feature pyramid networkÀÇ °ËÃâ °á°ú¿¡¼ º¹ÀâÇÑ ¹è°æÀ¸·Î ÀÎÇÑ ¿À°ËÃâÀ» È¿°úÀûÀ¸·Î °¨¼Ò½Ãų ¼ö ÀÖ´Â ¸ÖƼ ÇÁ·¹ÀÓ ±â¹Ý ÈÄó¸® ¹æ¹ýÀ» Á¦¾ÈÇÑ´Ù. Á¦¾ÈµÈ ÈÄó¸® ¹æ¹ýÀº ÀüÈÄ ÇÁ·¹ÀÓ¿¡¼ °ËÃâµÈ °´Ã¼ »çÀÌÀÇ À§Ä¡ Â÷ÀÌ ¹× Å©±â Â÷À̸¦ ºñ±³ÇÏ¿© ¿À°ËÃâ ¿©ºÎ¸¦ ÆÇ´ÜÇÑ´Ù. ¶ÇÇÑ µö·¯´× ³×Æ®¿öÅ©ÀÇ ÇнÀÀ» À§ÇÏ¿© Á÷Á¢ ÃÔ¿µÇÑ 360°³ÀÇ µå·Ð ¿µ»ó¿¡¼ È®º¸ÇÑ 44,986ÀåÀÇ ÁÖ¼®ÀÌ ´Þ¸° µå·Ð µ¥ÀÌÅÍ ¼¼Æ®¸¦ ±¸ÃàÇÏ¿´´Ù. Á¦¾È ÈÄó¸® ¹æ¹ýÀ» Àû¿ëÇÏ¿´À» °æ¿ì ¸ðµç Æò°¡ ½ÃÄö½ºÀÇ false positive°¡ 80% ÀÌ»ó °³¼±µÊ°ú µ¿½Ã¿¡ F-measureµµ Áõ°¡ÇÏ´Â °á°ú¸¦ º¸¿´´Ù.
|
¿µ¹®³»¿ë (English Abstract) |
With the rapid development in the field of drones, the need for a surveillance system to prevent accidents caused by drones has increased. Considering the high speed of drones, they must be detected from a distance. In long-distance images, the target size is very small and the background can be complex. Even if a deep learning technique is used for object detection the false detection rate remains very high. This paper introduces a multi-frame based post-processing method that can effectively reduce the false detection rate of feature pyramid network (FPN), which works well for tiny-object detection. The proposed post-processing method indexes detected objects and compares the distance and size difference concerning the corresponding objects of the previous frame to determine whether it is a false positive (FP) or not. FPN is trained on 44,986 images with annotations from 360 image sequences taken by hand. Experimental results show that the proposed method reduces the FP rate overall evaluation sequences by more than 80% and also increases the F-measure.
|
Å°¿öµå(Keyword) |
CNN
µö·¯´×
µå·Ð
FPN
°´Ã¼ ŽÁö
CNN
deep learning
drone
FPN
object detection
|
ÆÄÀÏ÷ºÎ |
PDF ´Ù¿î·Îµå
|