Visual Object Tracking
Embedded Files
1.Context-Aware Environmental Residual Correlation Filter for Robust Tracking
1.Context-Aware Environmental Residual Correlation Filter for Robust Tracking
Tracking objects in real-world scenarios requires handling rapid motion, occlusions, scale variations, and background clutter. In this project, I developed a correlation filter-based tracker (CAERDCF) that utilizes environmental residual learning and multi-feature fusion to improve robustness against challenging conditions. The method significantly enhances tracking adaptability by incorporating context-aware patches and a selective spatial regularizer, preventing drift and boundary artifacts.
Skills Learned:
Skills Learned:
Correlation filter-based tracking and spatial regularization techniques
Feature extraction from deep convolutional networks
Benchmarking on datasets like OTB2015, UAV123, LaSOT, and GOT10K
Key Highlights
Key Highlights
Increased tracking precision by 12.9% on OTB2015 and 16.1% on TempleColor128 compared to BACF.
Introduced an efficient environmental data acquisition strategy without additional computational cost.
Published in Mathematics (MDPI).
Overall framework of Proposed method
Overall framework of Proposed method
2.Multi-Level Hybrid Siamese Network with Box Adaptive Approach (SiamAdapt)
2.Multi-Level Hybrid Siamese Network with Box Adaptive Approach (SiamAdapt)
Siamese-based trackers have demonstrated strong performance, but many rely on predefined anchor boxes, making them sensitive to target scale variations. This project introduced SiamAdapt, a proposal-free tracking approach that eliminates anchor dependency, using a box-adaptive method to improve tracking accuracy. The framework incorporates ranking-based classification loss and IoU-guided ranking loss, optimizing both localization and classification.
Skills Learned:
Skills Learned:
Siamese-based object tracking and feature matching techniques
Designing anchor-free object detection methods
Optimization of classification and regression tasks in tracking
Key Highlights
Key Highlights
Achieved leading performance on OTB100, UAV123, GOT-10K, VOT2018, VOT2019, and LaSOT.
Operates at real-time speed with high efficiency.
Published in Multimedia Tools and Applications (Springer).
Overall framework of proposed method
Overall framework of proposed method
Overall working flow of Siamadapt
Comparison Results
Comparison Results
Illustration of trcking robustness under various challenging conditions
3.Keypoint Prediction Enhanced Siamese Tracker with Attention Mechanisms
3.Keypoint Prediction Enhanced Siamese Tracker with Attention Mechanisms
This project aimed to improve localization in tracking by incorporating keypoint prediction within a Siamese network. The model refines tracking accuracy using a progressive heatmap refinement technique that sharpens target positions. Additionally, self-attention and cross-attention mechanisms are used to enhance feature representations, improving tracking robustness in cluttered scenes.
Skills Learned:
Skills Learned:
Keypoint-based tracking and heatmap refinement
Self-attention and cross-attention techniques for feature extraction
Handling occlusion, scale variation, and appearance changes in tracking
Key Highlights
Key Highlights
Improved accuracy on LaSOT, GOT-10K, OTB100, and UAV123, surpassing state-of-the-art methods.
Introduced a multi-stage refinement process for precise object localization.
Published in Expert Systems with Applications (Elsevier).
Comparison Results
Comparison Results
Overall working flow of proposed SiamRAKPN approach which includes both self and cross attention mechanisms along with the KPN network.
Comparison Results
Comparison Results
Precision and success plots on UAV123, OTB100 and LaSOT datasets.
Page updated
Google Sites
Report abuse