One-Stage Anchor-Free Online Multiple Target Tracking With Deformable Local Attention and Task-Aware Prediction

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• The tracking-by-detection paradigm is standard for multiple target tracking.
• Sequential execution of target detection, appearance feature embedding, and data association limits tracking efficiency.

Purpose of the Study:

• To develop a one-stage, anchor-free multiple task learning framework for enhanced multiple target tracking.
• To improve tracking speed and accuracy by parallelizing detection and feature embedding.

Main Methods:

• A one-stage anchor-free framework processing target detection and appearance feature embedding in parallel using shared feature maps.
• Introduction of a deformable local attention module for discriminative feature extraction.
• Implementation of a task-aware prediction module with deformable convolutions for optimized task execution.
• Development of regression range overlapping and sample reweighting training strategies for dense scenes.
• An appearance-enhanced non-maximum suppression method to mitigate over-suppression in crowded environments.

Main Results:

• The proposed framework significantly increases tracking speed by performing detection and feature embedding concurrently.
• The deformable local attention and task-aware prediction modules yield more discriminative features and accurate predictions.
• Novel training strategies and non-maximum suppression effectively handle challenges in dense and crowded scenes.
• The implemented online multiple target tracker demonstrates high accuracy alongside remarkable speed.

Conclusions:

• The proposed one-stage, anchor-free framework offers a substantial improvement in multiple target tracking efficiency.
• The novel modules and training strategies effectively address limitations of traditional methods, especially in complex scenarios.
• This research presents a highly effective and fast solution for real-time multiple target tracking applications.