Multi-object tracking (MOT) using vision sensors remains a challenging problem, particularly in dynamic backgrounds and severe occlusions. Existing methods, relying on holistic appearance or spatial cues, fail to capture detailed information within object regions and the background, resulting in inaccurate and inconsistent tracking. To address these issues, we propose a novel Detail-Driven Multi-Object Tracking (DD-MOT) method that revisits and leverages fine-grained cues to discover the dynamics of both object regions and the background, facilitating the recovery and association of trajectories. Specifically, the proposed method consists of three key modules: (i) Points Trajectories Generator, (ii) Camera Motion and Occlusion Compensation, and (iii) Fine- and Coarsegrained Association. The Points Trajectories Generator is responsible for generating fine-grained cues by sampling an initial set of points, generating point trajectories, and refining the initial set of points. The Camera Motion and Occlusion Compensation module utilizes the background and object point trajectories to correct background motion and recover occluded object bounding boxes. Finally, the Fine- and Coarse-grained Association module leverages point trajectory cues to assist in establishing a more effective association strategy, combining both fine-grained and coarse-grained spatial cues from the object bounding boxes. We evaluate our DD-MOT method on several benchmark datasets, including MOT17, MOT20, demonstrating that it consistently outperforms state-of-the-art (SOTA) methods in key metrics such as HOTA (64.2, 62.6), MOTA (80.8, 78.1), and IDF1 (78.8, 77.1).

Deep neural networks for image denoising typically follow an encoder–decoder model, with convolutional (Conv) layers as essential components. Conv layers apply zero padding at the borders of input data to maintain consistent output dimensions. However, zero padding introduces ring-like artifacts at the borders of output images, referred to as border effects, which negatively impact the network’s ability to learn effective features. In traditional methods, these border effects, associated with convolutional/deconvolutional operations, have been mitigated using patch-based techniques. Inspired by this observation, patch-wise denoising algorithms were explored to derive a CNN architecture that avoids border effects. Specifically, we extend the patch-wise autoencoder to learn image mappings through patch extraction and patch-averaging operations, demonstrating that the patch-wise autoencoder is equivalent to a specific convolutional neural network (CNN) architecture, resulting in a novel residual block. This new residual block includes a mask that enhances the CNN’s ability to learn border features and eliminates border artifacts, referred to as the Border-Enhanced Residual Block (BERBlock). By stacking BERBlocks, we constructed a U-Net denoiser (BERUNet). Experiments on public datasets demonstrate that the proposed BERUNet achieves outstanding performance. The proposed network architecture is built on rigorous mathematical derivations, making its working mechanism highly interpretable. The code and all pretrained models are publicly available.