Hyperspectral videos (HSVs) contain abundant spectral information to facilitate the capture of distinctive spectral characteristics of the target. In RGB images, traditional tracking algorithms are prone to failure when confronted with targets that share similar shape, size, or color with the background, or low spatial resolution. Hyperspectral images provide detailed information about the internal structure and chemical composition of the target in the form of a three-dimensional data cube, where each target possesses a unique spectral curve. However, as the number of bands increases in hyperspectral images, both data complexity and computational complexity escalate, with diminishing data processing efficiency. Therefore, effective data compression becomes crucial. The occlusion problem frequently affects tracking accuracy and impedes real-time tracking implementation of target tracking tasks. Consequently, we aim to address challenges related to data processing and occlusion in hyperspectral target tracking by providing an efficient algorithm for reducing spectral matching discrepancies and suppressing tracking drift.

The algorithm is based on the context filter framework and incorporates the scale filter from the DSST algorithm as the scale estimation module. By computing the structure tensors of both the target and search regions, we extract edge structure features, reconstruct their respective structure tensors, and decompose them to obtain feature roots and corresponding feature vectors. By calculating the Mahalanobis distance between the target region and background region, we derive a multi-dimensional spectral weight which is then multiplied with the structure tensor of the search region. Finally, we calculate the Euclidean distance to achieve dimensionality reduction to bring about an image that is copied into three channels and inputted into the VGG19 network for extracting depth features. These features are subsequently fed into an enhanced context filter which improves upon traditional methods by enhancing cyclic negative sample collection techniques. By calculating each sample's interference factor, we select only the top four samples for training purposes to obtain response graphs for current frames. Based on the calculated average peak correlated energy (APCE) score of current frames, a decision is made on whether to fuse the initial frame's response graph to suppress tracking drift. Due to the propensity of the one-way learning mode in correlation filtering to introduce background noise leading to model errors over time resulting in tracking drift, accumulated errors should be minimized.

To verify the effectiveness of the proposed algorithm, we select four hyperspectral target tracking algorithms and compare them in the experiment. Meanwhile, a specific sequence is selected on the test set to visualize the performance of the proposed algorithm compared with the other four algorithms. Figure 4 shows the qualitative analysis results of various algorithms in selected sequences. In the ball sequence, the ball is moved and blocked by the finger, rolling back and forth. Since the proposed algorithm has improved the sampling method of the background negative sample, it can be stably tracked. In the toy sequence, two toys move alternately with each other, and the target toy is disturbed by another analogue toy. The proposed algorithm adaptively updates and adjusts the target model by adopting the initial model of the first frame to achieve tracking robustness. We evaluate the algorithms from two aspects of tracking accuracy and success rate. Tables 1 and 2 show the accuracy and success rate of the five algorithms respectively. Figure 5 shows the accuracy and success rate curves of each algorithm on the test sequence. Figures 6 and 7 demonstrate the accuracy and success rates associated with target occlusion and fast-moving challenges. As shown in Fig. 5, the proposed algorithm ranks first in terms of accuracy and success rate on the total test sequence. Specifically, the accuracy increases by 4.1% and the success rate grows by 4.5% compared to SiamBAG. Due to the utilization of adaptive tracking regression modules, the algorithm has strong robustness. As shown in Fig. 6, in the case of target occlusion, the accuracy of the proposed algorithm is only 0.9% which is higher than that of the second place, and the success rate is 0.4% higher, which is because the multi-feature fusion strategy is not employed. Additionally, as shown in Fig. 7, under the challenge of fast-moving targets, the accuracy of the proposed algorithm is 1.4% which is higher than that of the second place, and the success rate is 7.1% higher, with excellent adaptability shown. Table 3 presents the accuracy and success rate of the ablation experiment and reveals that the proposed method improves the algorithm robustness.

The selection of positive and negative samples is improved in the context filter framework and a hyperspectral target tracking algorithm based on structure tensor reduction and improved context filter is proposed. Texture information of the target is extracted using structure tensors, and multi-band spectral information is combined to conduct dimensionality reduction pre-processing of the image. Spectral information is introduced to the positive samples of the target, and the negative samples are screened, with the samples with the strongest interference factors selected for training. The experiments show that the proposed SI-HVT algorithm has good tracking ability in the aspects of occlusion resistance and fast movement. In future work, we will improve the sampling method of the filter to divide the negative samples more carefully and collect the positive samples not limited to the current frame. Additionally, we will try to extract features in a diversified manner. The multi-feature fusion strategy can make the algorithm better resistant to challenges such as light change and background clutter.