MMLT: Efficient object tracking through machine learning-based meta-learning
Bibek Das, Asfak Ali, Suvojit Acharjee, Jaroslav Frnda, Sheli Sinha Chaudhuri · Results in Engineering · 2025
Object tracking in computer vision is challenging due to the complexities of spatiotemporal data, including occlusions, abrupt motion changes, temporal coherence, object identity maintenance, and scale variations. While Deep learning algorithms address these challenges, however, they typically require significant computational resources, exhibit high complexity, and demand large amounts of training data. To mitigate these constraints, various strategies—such as lightweight neural networks and model compression techniques—have been developed. In contrast, traditional machine learning and classical computer vision methods like Kernelized Correlation Filters (KCF), Tracking, Learning, and Detection (TLD), and Bootstrap Aggregating (BOOSTING), lacks reliability in performance. This paper introduces a machine learning-based approach to one-shot meta-learning for more efficient object tracking. The proposed hybrid model refines predictions from traditional tracking methods using machine learning to enhance performance. This method offers lower computational complexity, requires fewer resources, and needs minimal training data. The proposed model achieves a frame rate of 13.74 FPS, which, while below real-time performance, maintains a trade-off between accuracy and computational efficiency, making it suitable for applications with moderate latency tolerance. The meta-learning-based approach is evaluated on the VOT2017, OTB50, and GOT10K datasets, outperforming existing deep-learning models in most cases. On the OTB50 dataset, the model with XGBoost achieved an OTE of 9.12%, a precision of 92.0%, and a success rate of 86.0%. On the GOT10k dataset, the model with a Decision Tree meta-learner recorded an average overlap of 62.2%, with success rates of 57.9% at 0.5 IoU and 52.0% at 0.75 IoU. On the VoT dataset, the Decision Tree meta-learner attained 79.0% EAO, 88.0% accuracy, and a robustness score of 20.0%. • Combines CSRT, TLD, MIL, and Boosting with adaptive selection for real-time object tracking. • Utilizes meta-learning to enhance tracking accuracy in dynamic and complex environments. • XGBoost meta-learner achieved 92.0% precision and 86.0% success on the OTB-50 dataset with 9.12% OTE. • Decision Tree meta-learner hit 62.2% overlap and 57.9% success at 0.5 IoU on the GOT10K dataset. • On VOT2017, Decision Tree meta-learner reached 79.0% EAO and 88.0% precision.