SVMOT: Stationary Vehicle Multi-Object Tracking Dataset

SVMOT is a multi-object tracking (MOT) dataset designed for traffic scenes with a high proportion of stationary vehicles and frequent object occlusions. The dataset follows the MOTChallenge format and extends it with visibility annotations, pairwise occlusion annotations, depth ordering annotations, and a per-object stopped attribute. Compared with existing MOT benchmarks, SVMOT is specifically designed for long-term stationary vehicles, challenging occlusions, and identity preservation under prolonged interactions between moving and non-moving traffic participants. The dataset contains 30 video sequences collected from 6 unique traffic scenes, comprising 26,900 frames, 243,462 annotated bounding boxes, and 1,017 unique identities. All sequences are distributed in a single archive preserving the original dataset structure. The release also includes official benchmark split files and scene-level metadata. This dataset is intended for research on occlusion-aware tracking, long-term identity preservation, and robust tracking in urban traffic scenes containing many stationary or temporarily stopped vehicles. Because multiple sequences may originate from the same physical scene, the recommended benchmark split is defined at the scene level to avoid scene leakage between training and test subsets. Files included in this release: svmot_data_v1.0.zip: full dataset archive README_RELEASE_v1.0.md: release documentation train.txt: recommended training split test.txt: recommended test split by_scene.json: scene-to-sequence mapping and split metadata SHA256SUMS: checksums for release files The archive svmot_data_v1.0.zip contains the data directory with sequence subdirectories in MOTChallenge-like format, as well as classes.txt. Each sequence includes image frames, ground-truth annotations, detections, sequence metadata, ROI masks, and occlusion annotations. If you use this dataset in your research, please cite the corresponding Zenodo record.