Aug 2026· IEEE Sensors Journal· Vol 26, pp. 24708-24716· 0 citations
Abstract
Motion-only multiobject tracking (MOT) suffers from ID switches in uniform-appearance and deformation-heavy scenes. In these settings, appearance cues become less reliable, so stable identities depend mainly on motion information. Existing methods often process all bounding-box variables together, which can weaken cues needed for prediction and matching. We address this problem by treating each trajectory as a heterogeneous multivariate time series (MTS) and redesigning the motion-only pipeline for embedding, encoding, and matching. HSMTrack separates box-state variables before modeling their temporal and cross-variable relationships, then uses deformation-aware matching for identity association. The method requires no appearance branch and can serve as a post-detection motion module in vision-sensor tracking pipelines. Its SSM-based encoder has linear complexity with respect to trajectory length, reducing modeling cost compared with attention-based alternatives. HSMTrack achieves 59.6 IDF1 and 42.9 AssA on DanceTrack, and 77.9 IDF1 and 67.2 AssA on SportsMOT. Under a unified end-to-end protocol, it reaches 34.1 frames/s on RTX 4090 and 10.5 frames/s on Jetson Orin NX.
This work introduces a system that combines the strong sequential constraints of SLAM with the flexibility and global optimization of offline SfM, enabling the metric reconstruction of arbitrary, long, uncalibrated videos.
Zador Pataki, Paul-Edouard Sarlin, Marc Pollefeys· arXiv.org· 0 citations
Single-object tracking from unmanned aerial vehicles (UAVs) is complicated by small target size, rapid camera ego-motion, and frequent occlusion, all of which degrade the appearance cues that Transformer trackers rely on. We present HOMA-ST, which recovers a complementary motion signal by decomposing the short-term dynamics of the bounding box into three physically meaningful branches — translation, geometric variation, and trajectory — and extracting a 21-dimensional high-order motion descriptor in which every dimension has an explicit interpretation and a known minimum-frame requirement. The three branches are fused by a branch-token self-attention block, and the resulting motion energy is injected as a learnable, gated residual on the search-side queries at every cross-attention layer. An auxiliary motion-consistency loss further couples spatial accuracy to temporal plausibility. On UAV123, UAVDT, DTB70, and VisDrone-SOT, HOMA-ST improves success AUC over OSTrack by 1.7 to 2.1%, with the largest gains on fast-motion and occlusion subsets. A further evaluation on 20 live-flight sequences recorded on a DJI Tello EDU platform confirms a 4.5% AUC improvement under real deployment conditions. The results indicate that HOMA-ST provides an effective framework for robust UAV tracking in both benchmark and real-flight scenarios.
Peng Gu, Zhanlin Qiu, L. Cherikbayeva et al.· IEEE Access· 0 citations
In the UAV traffic monitoring scene, vehicle targets usually have the characteristics of small scale, dense distribution and complex camera motion. These factors will reduce the detection reliability and interfere with the data association process, which can easily lead to trajectory breakage and identity switching problems. Aiming at the above problems, this paper proposes a unified detection and tracking framework, termed GSC-ByteTrack (Global Motion, Scale-Adaptive, and Class- Consistency ByteTrack). Firstly, an improved YOLOv8-P2 detection model is introduced to enhance the high-resolution feature expression ability of small targets. On this basis, three modules of global motion compensation, scale adaptive threshold division and category consistency constraint are designed to alleviate the influence of camera motion, improve the stability of small target correlation and reduce cross-category mismatch. Experiments conducted on the VisDrone2019- MOT dataset demonstrate that the proposed method achieves superior tracking performance compared with existing approaches, reaching 53.4% HOTA and 69.4% IDF1. Compared with the YOLOv8n + ByteTrack baseline, the proposed framework improves HOTA and IDF1 by 3.0 and 7.1 percentage points, respectively, indicating substantial gains in overall tracking accuracy and identity consistency. The proposed framework effectively balances detection sensitivity and association stability, providing a robust and efficient solution for UAV multi-object tracking in complex traffic scenarios.
Airborne infrared small-object tracking is crucial for applications such as autonomous reconnaissance and border surveillance. Unlike visible-light imagery, infrared data provides stable imaging in low-light and hazy conditions. However, tracking in this domain is exceptionally challenging due to the diminutive size of targets (approximately $13\times 23$ pixels), their weak textural features, and high susceptibility to complex background interference, including heat sources and sensor noise. Existing re-identification (ReID) models, predominantly designed for visible-light data, struggle to extract robust and discriminative features from such impoverished inputs. To address these limitations, we introduce a dual-branch joint decision tracking framework that synergistically integrates motion prediction with hierarchical appearance representation. The framework’s core is a hierarchical dual-branch ReID model: its shallow branch extracts stable intensity and structural features for short-term consistency, while its deep branch captures high-level semantic features for robust re-identification after occlusion. Furthermore, a novel time-window-based motion prediction module enhances trajectory accuracy by aggregating and statistically weighting multi-frame kinematic data. Crucially, we develop an adaptive joint decision mechanism that dynamically fuses these motion and hierarchical appearance cues. This mechanism constructs a unified cost matrix by scaling motion distances based on scene-level velocity statistics and selectively employs deep features to resolve ambiguities. Evaluated on a public infrared tracking dataset, our method demonstrates robust overall tracking performance, achieving a Multiple Object Tracking Accuracy (MOTA) of 95.6% and an Identification F1 (IDF1) score of 97.7%. These results underscore the framework’s enhanced stability and robustness in complex, dynamic environments, offering a more reliable solution for airborne infrared multi-object tracking than existing approaches.
Mingyu Hong, Xue Jin, Yuan Liu et al.· IEEE Access· 0 citations
Currently, prevalent object-tracking methods are mostly trained using image pairs (i.e. a template image and a search image), and this training paradigm makes it difficult to capture temporal correlations in the object’s continuous motion. Meanwhile, methods that depend on consecutive video frame training incur a drastic surge in hardware resource consumption. To overcome the above issues, an object tracking network model called TraTeTrack, based on historical motion trajectory modeling and multi-feature fusion, is proposed in this paper. Specifically, the visual features of the template and search images are extracted by the vision transformer module, and the TimesNet module is utilized to obtain the cross-frame temporal correlation features from historical motion trajectories. Then, the position of the object in the next frame is predicted through the detection head after the visual features and temporal features are fused. Moreover, a multi-strategy collaborative mechanism is designed to dynamically adjust the search range and update the object template, thereby effectively addressing challenges such as object occlusion. The core innovation of TraTeTrack is to introduce historical trajectory guidance into the conventional template-search image-pair tracking framework. Instead of replacing visual image pairs, the proposed trajectory branch complements them by modeling low-dimensional historical bounding-box sequences. To evaluate the tracking performance, the proposed TraTeTrack is compared with the state-of-the-art models on multiple datasets, and ablation experiments are conducted to validate the effectiveness of each module in the proposed model. Experimental results reflect that the proposed model achieves an area under curve of 70.5% on the LaSOT dataset with a frame rate of 90 FPS, which satisfies the requirements of real-time tracking.
Yong Tao, Haibin Wang, Jing-Lin Ma et al.· Measurement science and tech...· 0 citations
Optical flow forms a fundamental information for various motion related vision problems: e.g., SLAM, visual odometry, and object motion estimation. Event cameras are ideal vision sensors for on-line, dynamic tasks that require optical flow estimation, as they have high temporal resolution, high dynamic range and low latency. However, efficiently decoding optical flow from events for high frequency operation while maintaining accuracy is still an open problem. Batch-based optical flow algorithms (CNN or contrast maximisation) accumulate event data over a short period of time and achieve state-of-the-art performance in terms of accuracy, but at the cost of algorithm latency and lower update rates (on par with traditional cameras). In contrast, event-by-event algorithms only compute flow vectors in small, local regions, achieving a lower latency, but losing accuracy when global information is ignored. In this paper, we introduce a spatio-temporal registration framework to increase accuracy of current state-of-the-art event-by-event flow estimation, while also introducing a twofold algorithm acceleration approach and a real-time implementation strategy to mitigate the impact of computation scaling with event rate. We evaluate our event by-event optical flow algorithm on MVSEC, achieving state-of-the-art results for event-by-event algorithms, and performance comparable to batch-based methods. Our method is also computationally efficient, enabling processing of the higher resolution DSEC dataset, and is the only event-by-event algorithm tested to run completely in real time. Furthermore, we demonstrate its effectiveness and efficiency through qualitative evaluations on the ECD and the high-resolution M3ED datasets. Finally, we introduce a moving object dataset, which is outside the autonomous driving domain, to evaluate the general applicability of the proposed optical flow algorithm. The code is available open-source [CODE AVAILABLE ON ACCEPTANCE].
Zhichao Li, Arren J. Glover, Lorenzo Natale et al.· IEEE Transactions on Pattern...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
