Detection based long term tracking in correlation filter trackers

Feature Projection and SVR-Based Model Update for Object Tracking

Siamese network-based object tracking algorithms have recently gained popularity due to continuous improvement in tracking accuracy and robustness. However, these trackers are often limited by inefficient initialization and unstable online update. The model drift accumulates as more frames are processed. To address the above issues, this paper introduced a method that combines both deep learning and conventional regression algorithms (SVR, Support Vector Regression). Specifically, we propose a feature projection algorithm, which can effectively reduce the dimension of the features extracted from deep neural networks and improve the distinguishability of them at the same time. To make the features more robust for SVR training, we further propose two feature aggregation methods at both the channel level and the spatial level. We update the SVR model online instead of the deep neural network to make the tracking process more robust. Extensive experiments on four challenging benchmarks indicate that our proposed tracker is superior to baseline methods both qualitatively and quantitatively.

Aerial infrared target tracking method based on KCF for frequency-domain scale estimation

The kernel correlation filter (KCF) tracking algorithm encounters the issue of tracking accuracy degradation due to large changes in scale and rotation of aerial infrared targets. Therefore, this paper proposes a new scale estimation KCF-based aerial infrared target tracking method, which can extract scale feature information of images in the frequency domain based on the distribution characteristics and change laws of frequency-domain energy. In addition, the proposed method can improve the accuracy of target scale information estimation. First, the KCF tracking algorithm is used to obtain the target position. Then, spectral eigenvalues are calculated as eigenvectors, and frequency-domain rotation scale invariance is adopted to extract the eigenvector between two frames as the target rotation change information. Reverse rotation is performed on the current frame spectrum map for isolating the effects of target rotation on scale information estimation. Then, the current target scale is estimated on the basis of the eigenvectors between the adjacent frames. Finally, the length-to-width ratio and the scale of the tracking box are updated on the basis of the target rotation information, which improves the adaptability of the tracking box to changes in the target scale and rotation. The results indicate that the proposed algorithm is suitable for stable tracking of target scales and rapid changes in attitudes. The average tracking accuracy and the average success rate of the algorithm are 0.954 and 0.782, which represent improvements of 5.3% and 18.9%, respectively, compared with the KCF algorithm. The average tracking success rate is improved by 4.1% compared with the discriminative scale space tracker algorithm, and the average tracking performance is better than that of related filter tracking algorithms based on other scale estimation methods.