Object Tracking with Channel Group Regularization and Smooth Constraints Using Improved Dynamic Convolution Kernels in ITS

Aiming at the problem that the correlation between multi-channel feature representation and filter structure is not considered in the objective function modeling, a object tracking algorithm with channel group regularization and time series smooth constraint using improved dynamic convolution kernel...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Multimedia tools and applications Jg. 84; H. 29; S. 35191 - 35215
Hauptverfasser: Sun, Jinping, Li, Dan
Format: Journal Article
Sprache:Englisch
Veröffentlicht: New York Springer US 01.09.2025
Springer Nature B.V
Schlagworte:
1231: IoT-driven Computer Vision Technology for Smart Transportation Applications
Accuracy
Algorithms
Artificial neural networks
Channels
Computer Communication Networks
Computer Science
Constraints
Correlation
Data Structures and Information Theory
Feature maps
Intelligent transportation systems
Localization
Multimedia Information Systems
Neural networks
Object recognition
Process controls
Regularization
Spatial data
Special Purpose and Application-Based Systems
Task complexity
Tracking
Vision systems
Time series smoothing
Sparse representation
Correlation filter
Object tracking
Dynamic convolution kernel
ISSN:1573-7721, 1380-7501, 1573-7721
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Aiming at the problem that the correlation between multi-channel feature representation and filter structure is not considered in the objective function modeling, a object tracking algorithm with channel group regularization and time series smooth constraint using improved dynamic convolution kernels is proposed. Firstly, the elements in the filter are grouped using spatial and channel properties, the time-domain correlation of the object model is constrained by the low-rank kernel norm. A group regularization term is constructed to describe the correlation between channels with a mixed-norm structured sparsity constraint to learn lower-dimensional filters. Then, after extracting some channels in the feature map, part of the spatial information of each channel is retained to generate an efficient dynamic convolution kernel through the channel information de-redundant attention mechanism to obtain an optimized lightweight convolutional neural network. Finally, combining the advantages of hand-crafted features and deep convolutional features, the complementary localization of the object from coarse to fine is realized with the help of the constructed efficient feature model and the learned filter. The experimental results on public datasets show that the proposed algorithm can adapt to the tracking tasks of various complex traffic scenes, and enhance the tracking performance of the existing models. The proposed algorithm improves the discriminative property of the object model and the self-adaptability of the spatio-temporal information of the dynamic convolution kernel, and can be applied to neuromorphic vision system of intelligent transportation systems.
Bibliographie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:1573-7721
1380-7501
1573-7721
DOI:10.1007/s11042-022-14294-w