ClarityTrack for multi object tracking via hierarchical association and environment specific cost matching

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Existing multi-object tracking (MOT) methods often use fixed ratios for motion and appearance cues.
• This approach is suboptimal as cue reliability varies with scene conditions, leading to tracking errors.

Purpose of the Study:

• To develop an adaptive MOT system that addresses limitations of fixed-weight fusion.
• To enhance tracking accuracy and robustness across diverse environmental conditions.

Main Methods:

• Proposed ClarityTrack, an environment-aware, rule-based MOT system.
• Implemented Balanced Cascade Association (detection confidence-based, two-stage fusion).
• Utilized Condition-Aware Matching with Weights and Motion-Appearance Consistency Check for adaptive cue weighting.

Main Results:

• ClarityTrack achieved competitive HOTA, IDF1, and AssA scores on MOT17, MOT20, and DanceTrack datasets.
• Demonstrated superior performance over fixed-weight fusion methods.
• Showcased improved tracking quality and interpretability.

Conclusions:

• ClarityTrack offers an effective solution for adaptive multi-object tracking.
• The system's explicit, rule-based framework enhances interpretability and robustness.
• Pre-optimizing and conditionally switching parameters based on environment and track-detection quality is key.