The problem — Cell tracking sounds like detection with a subscript. It is not. Linking detections across frames requires resolving ambiguous associations — cells that divide, briefly disappear, or pass through occlusion — and a single dropped link breaks a complete lineage tree. A track that looks continuous in a video may still carry silent identity swaps that invalidate every downstream feature.
What it is / how it works — TrackMate (Tinevez et al., Methods 2017; v7 Ershov et al., Nature Methods 2022) is a modular Fiji plugin that separates detection from linking. The detection layer supports LoG/DoG blob detectors, morphological filters, and — in v7 — deep-learning segmentation models (Cellpose, StarDist, others) as drop-in detectors. The linking layer offers LAP (Linear Assignment Problem) solvers with configurable cost matrices for gap closing, split, and merge events. An interactive GUI overlays tracks on the raw image for immediate quality inspection; all steps are accessible via the ImageJ macro language for scripted or HPC execution. TrackMate is the de facto reference implementation used by the Cell Tracking Challenge, making its output directly comparable to published benchmarks.
Where it fits in the Fovea workflow — TrackMate is the tracking stage in light-sheet and live-cell pipelines, consuming fused volumes from PetaKit5D (after stitching by BigStitcher) and returning track tables that feed feature extraction and readout. For crowded or volumetrically complex acquisitions where segmentation uncertainty is high, ultrack is the preferred alternative — TrackMate shines in sparser scenes where its GUI-assisted parameter tuning is an advantage.
Where it breaks — LAP-based linking degrades rapidly as cell density increases and as the imaging interval grows relative to cell displacement — Seiffarth et al. (2024) quantify exactly this experiment-dependent collapse in tracking performance. TrackMate's default cost matrix has no mechanism for handling segmentation ambiguity (multiple competing masks); when the segmentation quality is uneven, link errors compound across time.