Spatial-Semantic and Temporal Attention Mechanism-Based Online Multi-Object Tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Multi-object tracking (MOT) is vital for many applications but faces challenges like occlusion, complex backgrounds, and real-time demands.
• Current tracking-by-detection methods are often computationally intensive due to sliding windows or anchoring schemes.

Purpose of the Study:

• To develop a more efficient and effective spatial-temporal attention scheme for multi-object tracking.
• To address challenges in candidate insertion and location, and improve occlusion estimation and classification.

Main Methods:

• Implemented a semantic-feature-based spatial attention mechanism and a novel Motion Model.
• Utilized online-learned target-specific convolutional neural networks (CNNs) for appearance model adaptation.
• Employed a temporal attention mechanism to balance current and historical frame information for online module updates.

Main Results:

• The proposed method demonstrated outstanding tracking performance on the KITTI and ATTD benchmarks.
• Achieved superior results in handling occlusions and target insertions compared to existing approaches.
• Met the stringent real-time requirements for practical applications.

Conclusions:

• The developed spatial-temporal attention scheme offers a significant advancement in multi-object tracking.
• The method provides an efficient and effective solution for complex tracking scenarios.
• Validated through extensive experiments on diverse datasets, confirming its practical applicability.