Tracking by detection has been the prevailing paradigm in the field of Multi-object Tracking (MOT). These methods typically rely on the Kalman Filter to estimate the future locations of objects, assuming linear object motion. However, they fall short when tracking objects exhibiting nonlinear and diverse motion in scenarios like dancing and sports. In addition, there has been limited focus on utilizing learning-based motion predictors in MOT. To address these challenges, we resort to exploring data-driven motion prediction methods. Inspired by the great expectation of state space models (SSMs), such as Mamba, in long-term sequence modeling with near-linear complexity, we introduce a Mamba-based motion model named Mamba moTion Predictor (MTP). MTP is designed to model the complex motion patterns of objects like dancers and athletes. Specifically, MTP takes the spatial-temporal location dynamics of objects as input, captures the motion pattern using a bi-Mamba encoding layer, and predicts the next motion. In real-world scenarios, objects may be missed due to occlusion or motion blur, leading to premature termination of their trajectories. To tackle this challenge, we further expand the application of MTP. We employ it in an autoregressive way to compensate for missing observations by utilizing its own predictions as inputs, thereby contributing to more consistent trajectories. Our proposed tracker, MambaTrack, demonstrates advanced performance on benchmarks such as Dancetrack and SportsMOT, which are characterized by complex motion and severe occlusion.

Multi-object tracking in unmanned aerial vehicle (UAV) videos is a critical visual perception task with numerous applications. However, existing multi-object tracking methods, when directly applied to UAV scenarios, face significant challenges in maintaining robust tracking due to factors such as motion blur and small object sizes. Additionally, existing UAV methods tend to underutilize crucial information from the temporal and spatial dimensions. To address these issues, on the one hand, we propose a temporal feature aggregation module (TFAM), which effectively combines temporal contexts to obtain rich feature response maps in dynamic motion scenes to enhance the detection capability of the proposed tracker. On the other hand, we introduce a topology-integrated embedding module (TIEM) that captures the topological relationships between objects and their surrounding environment globally and sparsely, thereby integrating spatial layout information. The proposed TIEM significantly enhances the discriminative power of object embedding features, resulting in more precise data association. By integrating these two carefully designed modules into a one-stage online MOT system, we construct a robust UAV tracker. Compared to the baseline approach, the proposed model demonstrates significant improvements in MOTA on two UAV multi-object tracking benchmarks, namely VisDrone2019 and UAVDT. Specifically, the proposed model achieves a 2.2% improvement in MOTA on the VisDrone2019 benchmark and a 2.5% improvement on the UAVDT benchmark.

Recently, advances in detection and re-identification techniques have significantly boosted tracking-by-detection-based multi-pedestrian tracking (MPT) methods and made MPT a great success in most easy scenes. Several very recent works point out that the two-step scheme of first detection and then tracking is problematic and propose using the bounding box regression head of an object detector to realize data association. In this tracking-by-regression paradigm, the regressor directly predicts each pedestrian’s location in the current frame according to its previous position. However, when the scene is crowded and pedestrians are close to each other, the small and partially occluded targets are easily missed. In this paper, we follow this pattern and design a hierarchical association strategy to obtain better performance in crowded scenes. To be specific, at the first association, the regressor is used to estimate the positions of obvious pedestrians. At the second association, we employ a history-aware mask to filter out the already occupied regions implicitly and look carefully at the remaining regions to find out the ignored pedestrians during the first association. We integrate the hierarchical association in a learning framework and directly infer the occluded and small pedestrians in an end-to-end way. We conduct extensive pedestrian tracking experiments on three public pedestrian tracking benchmarks from less crowded to crowded scenes, demonstrating the proposed strategy’s effectiveness in crowded scenes.