Trackers

Trackers

FlashTrack provides three tracking algorithms.

ByteTracker

Two-stage IoU-based tracker from ByteTrack (Zhang et al., 2022).

• Stage 1: Match high-confidence detections to tracks via IoU
• Stage 2: Match remaining tracks to low-confidence detections

from flashtrack import ByteTracker

tracker = ByteTracker(
    track_thresh=0.5,    # High/low confidence split
    track_buffer=30,     # Frames to keep lost tracks
    match_thresh=0.8,    # IoU threshold
)
tracks = tracker.update(detections)  # [N, 5] -> [M, 7]

Best for: Speed-critical applications, when ReID features are not available.

SORTTracker

Classic SORT algorithm (Bewley et al., 2016): Kalman filter + Hungarian + IoU.

from flashtrack import SORTTracker

tracker = SORTTracker(
    max_age=30,         # Max frames without update
    min_hits=3,         # Min hits for confirmation
    iou_threshold=0.3,  # IoU matching threshold
)
tracks = tracker.update(detections)

Best for: Simple tracking scenarios with minimal occlusion.

DeepSORTTracker

Deep SORT (Wojke et al., 2017): SORT + ReID cosine distance matching.

from flashtrack import DeepSORTTracker

tracker = DeepSORTTracker(
    max_age=70,
    max_cosine_distance=0.4,
    gallery_size=100,
)
tracks = tracker.update(detections, features=reid_embeddings)

Best for: Crowded scenes, frequent occlusions, re-identification after disappearance.

Output Format

All trackers return [M, 7] arrays:

[x1, y1, x2, y2, track_id, class_id, score]

Comparison

Tracker	ReID	Occlusion	Speed	Use Case
ByteTracker	No	Medium	Fastest	General purpose
SORTTracker	No	Low	Fast	Simple scenes
DeepSORTTracker	Yes	High	Moderate	Crowded scenes