Large Margin Object Tracking with Circulant Feature Maps (1703.05020v2)

Abstract: Structured output support vector machine (SVM) based tracking algorithms have shown favorable performance recently. Nonetheless, the time-consuming candidate sampling and complex optimization limit their real-time applications. In this paper, we propose a novel large margin object tracking method which absorbs the strong discriminative ability from structured output SVM and speeds up by the correlation filter algorithm significantly. Secondly, a multimodal target detection technique is proposed to improve the target localization precision and prevent model drift introduced by similar objects or background noise. Thirdly, we exploit the feedback from high-confidence tracking results to avoid the model corruption problem. We implement two versions of the proposed tracker with the representations from both conventional hand-crafted and deep convolution neural networks (CNNs) based features to validate the strong compatibility of the algorithm. The experimental results demonstrate that the proposed tracker performs superiorly against several state-of-the-art algorithms on the challenging benchmark sequences while runs at speed in excess of 80 frames per second. The source code and experimental results will be made publicly available.

• The paper presents LMCF, a novel tracking method that integrates structured SVMs with FFT-based correlation filters for robust, real-time detection.
• It employs multimodal target detection and a high-confidence model update strategy to mitigate issues like occlusions and rapid motions.
• Experiments on OTB-13 and OTB-15 show that LMCF outperforms state-of-the-art trackers, achieving accuracy and speeds exceeding 80 FPS.

Large Margin Object Tracking with Circulant Feature Maps

The paper "Large Margin Object Tracking with Circulant Feature Maps" by Mengmeng Wang, Yong Liu, and Zeyi Huang introduces an innovative approach to object tracking by leveraging structured SVMs and the computational efficiency of correlation filters. The methodology targets short-term, single-object visual tracking, focusing on real-time performance despite challenging scenarios such as occlusions and rapid motions.

Methodology

The authors propose a novel tracker termed Large Margin Object Tracking with Circulant Feature Maps (LMCF). The essence of this approach lies in its ability to combine the discriminative prowess of structured output SVMs with the speed advantages offered by correlation filters. This is principally achieved by employing dense circular samples in both training and detection phases, facilitated by the Fast Fourier Transform (FFT) to expedite computation.

Key elements of the methodology include:

• Multimodal Target Detection: The method improves target localization precision and mitigates model drift risks caused by similar objects or background noise. This enhancement allows for the reevaluation of target positions using multiple peaks in response maps.
• High-Confidence Model Update Strategy: This strategy circumvents model corruption by selectively updating only when detection confidence is high, assessed through peak response and APCE measures.
• Compatibility with CNN Features: Two tracker versions are implemented, using both conventional handcrafted features and CNN-based representations, thereby demonstrating the scalability and versatility of the approach.

Results

Experiments conducted on the OTB-13 and OTB-15 datasets reveal that LMCF outperforms several state-of-the-art algorithms in terms of accuracy and robustness, with a substantial speed advantage exceeding 80 FPS. Remarkably, LMCF's precision and success rates eclipse those of notable algorithms, including Struck, DSST, and others, while remaining computationally efficient.

The results also highlight the advantages of integrating CNN features, evidenced by the superior performance of DeepLMCF, albeit with a trade-off in processing speed. This demonstrates the method's adaptability to diverse feature types and its potential integration with advanced feature extraction techniques.

Implications and Future Directions

The proposed LMCF methodology presents significant practical implications for real-time visual tracking in applications ranging from surveillance to autonomous systems. The blend of structured SVM and fast correlation filtering paves the way for further exploration of hybrid models that combine high discriminative power with computational efficiency.

The adoption of CNN-based features, as shown in DeepLMCF, underscores the potential for even greater tracking accuracy, suggesting a promising trajectory for incorporating more advanced neural network architectures in object tracking paradigms.

Looking forward, future work may delve into enhancing the robustness of multimodal detection techniques or further optimizing the balance of precision and speed to tackle even more dynamic environments. Additionally, exploring the application of this methodology in multi-object tracking scenarios could extend its applicability in complex real-world settings.

In summary, the research provides a compelling advancement in real-time object tracking, presenting a scalable, efficient, and accurate solution that holds potential for broad application within the field of computer vision.