Improving Object Tracking by Added Noise and Channel Attention.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Convolutional Neural Network (CNN)-based trackers, particularly Siamese networks, offer good performance and efficiency.
• A key challenge in Siamese trackers is learning a generalized object model from large datasets, often leading to overfitting.

Purpose of the Study:

• To improve the generalization capability of Siamese trackers.
• To mitigate the overfitting problem in training generic object models.
• To enhance target localization and overall tracking performance.

Main Methods:

• Introduced input noise as a regularization technique during training data preparation.
• Implemented offline learning with additive noise for input data augmentation.
• Proposed feature fusion from both noisy and clean input channels.
• Integrated channel attention mechanisms to identify more discriminative target features.

Main Results:

• The proposed Input-Regularized Channel Attentional Siamese (IRCA-Siam) tracker demonstrated superior generalization over state-of-the-art methods.
• Feature fusion from noisy and clean inputs improved target localization accuracy.
• Channel attention further boosted performance by focusing on relevant target features.
• IRCA-Siam showed enhanced discrimination between target and background, improving fault tolerance.

Conclusions:

• Input noise regularization is an effective strategy for improving Siamese tracker generalization.
• The IRCA-Siam tracker offers enhanced performance, robustness, and generalization capabilities.
• Experimental results on multiple benchmark datasets validate the superiority of IRCA-Siam.