​
Overview

ByteTrack: Multi-Object Tracking by Associating Every Detection Box This module implements the ByteTrack algorithm for multi-object tracking, which achieves high performance by associating both high and low confidence detections. ByteTrack uses a multi-stage matching process that recovers true objects from low-confidence detections while filtering out background noise, making it robust for real-world tracking scenarios. ByteTrack associates every detection box instead of only high-confidence ones, utilizing similarities with existing tracklets to recover true objects from low-confidence detections while filtering out background. Uses a two-stage matching process: first matching high-confidence detections to existing tracks, then matching low-confidence detections to remaining unmatched tracks. This approach significantly improves tracking performance in challenging scenarios with occlusions and detection noise.

​
Class Overview

The ByteTracker class provides structured data management for bytetracker operations.

​
Parameters

​
track_activation_threshold
float
required
Detection confidence threshold for track activation. Range: [0.0, 1.0]. Default is 0.25 (25% confidence).
​
lost_track_buffer
int
required
Number of frames to buffer when a track is lost before removal. Range: [1, 100]. Default is 30 frames (~1 second at 30fps).
​
minimum_matching_threshold
float
required
IoU threshold for first-stage matching with high confidence detections. Range: [0.0, 1.0]. Default is 0.7 (70% overlap required).
​
minimum_consecutive_frames
int
required
Minimum consecutive frames before considering a track valid. Range: [1, 10]. Default is 3 frames.
​
second_match_threshold
float
required
IoU threshold for second-stage matching with low confidence detections. Range: [0.0, 1.0]. Default is 0.5 (50% overlap required).
​
assignment_threshold
float
required
IoU threshold for assigning tracker IDs to final detections. Range: [0.0, 1.0]. Default is 0.3 (30% overlap required).

​
Examples

import cv2
import pixelflow as pf
from ultralytics import YOLO

# Initialize tracker with default settings
tracker = pf.tracker.ByteTracker()
model = YOLO("yolo11n.pt")
video_path = "path/to/video.mp4"

# Process video frames
for frame in pf.video.get_video_frames(video_path):

​
Error Handling

This function may raise the following exceptions:
  • ValueError: If any threshold parameter is outside valid range [0.0, 1.0]
  • ValueError: If frame or buffer parameters are not positive integers

​
Notes

  • Uses Kalman filtering for motion prediction
  • Implements two-stage association: high-confidence then low-confidence
  • Automatically handles track state transitions (tracked → lost → removed)
  • Maintains track continuity through temporary occlusions
  • Filters out short-lived false positive tracks
  • All threshold parameters are automatically clamped to valid ranges
  • Optimized for real-time video processing (>30 FPS on modern hardware)
  • Memory usage scales linearly with number of active tracks
  • IoU computation is the primary bottleneck for scenes with many detections
  • Efficient track management prevents memory leaks in long videos