What is Multi-Object Tracking?
Multi-object tracking (MOT) goes a step beyond detection. Detection asks where objects are in one frame; tracking asks who each object is across time. Every player, ball or vehicle gets a persistent identity that survives from frame to frame — the foundation for any analytics built on trajectories rather than snapshots.
Hassan has specialized in exactly this problem for sports. At ID Sports Ventures in Berlin, his tracking systems keep every athlete's ID intact through the chaos of live play — the hard case that decides whether a sports product works. Read the deep dive on ByteTrack multi-object tracking for the technical mechanics.
Detection + Association
Every tracker is two problems chained together. First, detection: finding each object each frame. Second, association: deciding which new detections continue which existing tracks. The second problem is where identity survives or dies.
Association blends motion prediction, appearance features and detection confidence. Kalman filters predict where a player should be next; appearance models re-identify them after occlusion; and score thresholds decide how aggressively to link or kill tracks. Get the fundamentals from YOLO object detection.
YOLO & ByteTrack in Practice
Hassan pairs YOLO for real-time detection with ByteTrack for association. ByteTrack's defining idea is that every detection matters: instead of throwing away low-confidence boxes, it associates them too, recovering objects that other trackers write off.
In a dense field of similar-looking players, that distinction is everything. Low-confidence boxes often mean a partially occluded athlete — exactly the case where a naive tracker drops the ID and the downstream analytics break. When appropriate, appearance features extend the pipeline, drawing on computer vision fundamentals.
Handling Severe Occlusion
Occlusion is the tracker's ultimate test: one player walks in front of another and the two boxes merge or one vanishes. Hassan's systems survive this through layered defenses:
- Motion prediction — Kalman filtering keeps a track alive when its detection flickers out.
- Low-score association — ByteTrack-style matching recovers partially hidden players.
- Appearance re-identification — feature embeddings re-link a player after full occlusion.
- Tuned for sports — thresholds calibrated on real match footage, not clean benchmarks.
Sports Applications
With stable identities, everything downstream becomes possible. Automated scoring attributes each event to the right player; drill assessment measures how individuals move through a designed exercise; possession, workload and team shape analysis all rest on knowing who is who.
These are the capabilities behind the sports analytics work Hassan ships — and the reason teams pair him with pose estimation to build complete athlete-analysis products. For the career picture, see Sports Analytics Engineer.
“Tracking is where detection becomes intelligence — keeping every player's identity alive through occlusion is what makes analytics trustworthy.” — Muhammad Hassan Gul