Centroid-based graph matching networks for planar object tracking

Recently, keypoint-based methods have received more attention on planar object tracking due to their abilities to deal with partial noises, such as occlusion and out-of-view. However, robust tracking is still a tricky problem in the case of fast movement, large transformation and motion blur. The ke...

Celý popis

Uložené v:
Podrobná bibliografia
Vydané v:Machine vision and applications Ročník 34; číslo 2; s. 31
Hlavní autori: Li, Kunpeng, Liu, He, Wang, Tao
Médium: Journal Article
Jazyk:English
Vydavateľské údaje: Berlin/Heidelberg Springer Berlin Heidelberg 01.03.2023
Springer Nature B.V
Predmet:
Algorithms
Blurring
Centroids
Communications Engineering
Computer Science
Datasets
Deep learning
Design
Graph matching
Image Processing and Computer Vision
Inliers (landforms)
Localization
Methods
Movement
Networks
Occlusion
Original Paper
Pattern Recognition
Perturbation
Template matching
Tracking
Transformations
Vision systems
Planar object tracking
Deep learning
Centroid
Pose estimation
Keypoint
Graph matching
ISSN:0932-8092, 1432-1769
On-line prístup:Získať plný text
Tagy: Pridať tag
Žiadne tagy, Buďte prvý, kto otaguje tento záznam!
Popis
Shrnutí:Recently, keypoint-based methods have received more attention on planar object tracking due to their abilities to deal with partial noises, such as occlusion and out-of-view. However, robust tracking is still a tricky problem in the case of fast movement, large transformation and motion blur. The key reason is that there are not enough matching inliers to reconstruct the homography in the presence of such perturbations. To this end, we propose a novel centroid-based graph matching networks (CGN), which consists of two components: centroid localization network (CLN) and graph matching network (GMN). In detail, the CLN reduces the search range of the tracker from the entire image to the target region by locating the centroid of the target. The CLN gives the initial guess of the position, which guarantees the proportion of inliers matching the template. Then, the keypoints in the template and the target region are modeled as two graphs connected by cross-edges, and their correspondences are established by the GMN. The GMN overcomes the impact of large transformation by exploiting the stability of the graph structure. Finally, the transformation from the template to the current frame is estimated from the matched keypoint pairs by the RANSAC algorithm. In addition, the number of labeled points in previous datasets for training matching models is too small to cope with complex transformations, so we synthesize a large-scale dataset with labels to train the GMN. Experimental results on POT-210, POIC and TMT datasets show that our proposed method outperforms the state-of-the-art baseline methods in general, with significant improvements on fast movement and motion blur.
Bibliografia:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:0932-8092
1432-1769
DOI:10.1007/s00138-023-01382-6