TY - GEN
T1 - Globally Optimal Object Tracking with Complementary Use of Single Shot Multibox Detector and Fully Convolutional Network
AU - Lee, Jinho
AU - Iwana, Brian Kenji
AU - Ide, Shouta
AU - Hayashi, Hideaki
AU - Uchida, Seiichi
N1 - Publisher Copyright:
© 2018, Springer International Publishing AG, part of Springer Nature.
PY - 2018
Y1 - 2018
N2 - Object tracking is one of the most important but still difficult tasks in computer vision and pattern recognition. The main difficulties in the tracking task are appearance variation of target objects and occlusion. To deal with those difficulties, we propose a object tracking method combining Single Shot Multibox Detector (SSD), Fully Convolutional Network (FCN) and Dynamic Programming (DP). SSD and FCN provide a probability value of the target object which allows for appearance variation within each category. DP provides a globally optimal tracking path even with severe occlusions. Through several experiments, we confirmed that their combination realized a robust object tracking method. Also, in contrast to traditional trackers, initial position and a template of the target do not need to be specified. We show that the proposed method has a higher performance than the traditional trackers in tracking various single objects through video frames.
AB - Object tracking is one of the most important but still difficult tasks in computer vision and pattern recognition. The main difficulties in the tracking task are appearance variation of target objects and occlusion. To deal with those difficulties, we propose a object tracking method combining Single Shot Multibox Detector (SSD), Fully Convolutional Network (FCN) and Dynamic Programming (DP). SSD and FCN provide a probability value of the target object which allows for appearance variation within each category. DP provides a globally optimal tracking path even with severe occlusions. Through several experiments, we confirmed that their combination realized a robust object tracking method. Also, in contrast to traditional trackers, initial position and a template of the target do not need to be specified. We show that the proposed method has a higher performance than the traditional trackers in tracking various single objects through video frames.
UR - http://www.scopus.com/inward/record.url?scp=85042474427&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85042474427&partnerID=8YFLogxK
U2 - 10.1007/978-3-319-75786-5_10
DO - 10.1007/978-3-319-75786-5_10
M3 - Conference contribution
AN - SCOPUS:85042474427
SN - 9783319757858
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 110
EP - 122
BT - Image and Video Technology - 8th Pacific-Rim Symposium, PSIVT 2017, Revised Selected Papers
A2 - Hitoshi, Carlos
A2 - Paul, Manoranjan
A2 - Huang, Qingming
PB - Springer Verlag
T2 - 8th Pacific-Rim Symposium on Image and Video Technology, PSIVT 2017
Y2 - 20 November 2017 through 24 November 2017
ER -