MKB: Multi-Kernel Bures Metric for Nighttime Aerial Tracking

In recent years, many advanced visual object tracking algorithms have achieved significant performance improvements in daytime scenes. However, when these algorithms are applied at night on unmanned aerial vehicles, they often exhibit poor tracking accuracy due to the domain shift. Therefore, designing an effective and stable domain adaptation tracking framework is in great demand. Besides, as the complexity of the tracking task scene and a large number of samples, it is difficult to accurately characterize the feature differences. Moreover, the method should avoid excessive manual parameters. To solve above three challenges, this study proposes a metric-based domain alignment framework for nighttime object tracking of unmanned aerial vehicles. It achieves transfer learning from daytime to nighttime by introducing a multi-kernel Bures metric (MKB). MKB quantifies the distribution distance by calculating the difference between two domain covariance operators in the latent space. MKB also uses a linear combination of a series of kernel functions to accommodate complex samples of different scenes and categories, enhancing the ability of the metric to represent various sequences in the latent space. Besides, we introduce an alternating update mechanism to optimize the weights of kernels avoiding manual parameters. Experimental results demonstrate that the proposed method has significant advantages in improving tracking accuracy and robustness at nighttime.