Multi-Agent Multi-View Collaborative Perception Based on Semi-Supervised Online Evolutive Learning.

Area of Science:

• Edge Intelligence
• Machine Learning
• Computer Vision

Background:

• Edge intelligence systems rely on sensing devices for real-time scene recognition, but fixed models suffer from performance degradation due to changing environments.
• Continuous, unassisted learning is crucial for edge models to adapt to dynamic perception scenes and maintain service stability within online evolutive learning (OEL) systems.

Purpose of the Study:

• To address the generalization degradation problem in edge perception models.
• To propose and evaluate a novel semi-supervised online evolutive learning method for multi-view agents.

Main Methods:

• Introduced a multi-view agent's collaborative perception (MACP) method leveraging semi-supervised learning (SSL).
• Initialized view models using self-supervised learning for differentiated feature extraction.
• Fused multi-view predictions on unlabeled data to generate pseudo-labels for SSL consistency regularization, enhanced by parameter constraints for discriminative independence.

Main Results:

• The proposed MACP method demonstrated superior convergence efficiency and performance compared to existing multi-model and single-model SSL approaches on various benchmarks.
• Experimental results validated the effectiveness of collaborative perception and pseudo-label generation in improving model adaptability.

Conclusions:

• MACP effectively enhances the stability and generalizability of edge perception models in time-varying environments.
• The method shows significant potential for practical applications in real-world perception scenarios requiring continuous adaptation.