ASDTracker: Adaptively Sparse Detection With Attention-Guided Refinement for Efficient Multi-Object Tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Tracking-by-Detection methods are effective for multi-object tracking (MOT) but computationally expensive.
• High computational load stems from detection and re-identification (ReID) processes.
• Real-time applications require efficient MOT solutions.

Purpose of the Study:

• To develop an efficient tracking-by-detection algorithm (ASDTracker) that maintains high performance.
• To reduce computational burden by adaptively managing detector usage.
• To enable real-time MOT for applications like unmanned surface vehicles.

Main Methods:

• ASDTracker employs adaptively sparse detection with attention-guided refinement.
• It dynamically assesses occlusion to decide when to use the expensive detector.
• Refinement network trained with noisy shadow labels for non-key frames and uses lightweight appearance features instead of heavy ReID networks.

Main Results:

• ASDTracker achieves competitive generalization and robustness across four benchmarks.
• Demonstrates favorable inference speed compared to existing methods.
• Successfully deployed on an unmanned surface vehicle with high accuracy and low latency.

Conclusions:

• ASDTracker offers an efficient and effective solution for real-time multi-object tracking.
• The adaptive approach successfully balances computational cost and tracking accuracy.
• The method shows promise for real-world robotic applications requiring low-latency perception.