Learnable Diffusion-Based Amplitude Feature Augmentation for Object Tracking in Intelligent Vehicles

Data augmentation is an effective approach to enhance generalization of object tracking models, which has been widely applied into intelligent transportation to handle the challenging factors, such as lighting changes from day to night. However, current object tracking methods primarily employ pixel-level data augmentation techniques. These techniques only generate a limited number of samples with some fixed styles, making them difficult to cope with the complex scenario changes in intelligent transportation. To this end, we propose a learnable data augmentation approach for object tracking that is able to continuously augment the image amplitude features via a diffusion model, yielding numerous style-rich image feature representations. Specifically, we first employ fast Fourier transform (FFT) to decompose the spatial domain features into frequency domain features, which consist of phase components and amplitude components. The phase components encode the semantic information while the amplitude components encode the scene-style information. Especially, to generate diverse scene style features, we introduce a learnable diffusion model. During the training phase, we gradually introduce corresponding perturbation noise to the magnitude features and then use a U-Net network to progressively remove noise from the normal distribution variables. This allows us to learn the data distribution of the target and capture its data semantic structural information. By continuously adding noise to and denoising amplitude information, we enhance the tracker's ability to handle complex and style-rich scenes. Experiments show that the proposed method obtains the best generalization and tracking performance on multiple public datasets.