Learning person-person interaction in collective activity recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Collective activity recognition is challenging due to the difficulty of distinguishing activities from isolated individual actions.
• Person-to-person interactions are crucial cues for understanding group behaviors.

Purpose of the Study:

• To propose a novel learning-based approach for modeling person-person interactions in collective activity recognition.
• To develop a method that computes class-specific interaction patterns without relying on handcrafted features.

Main Methods:

• Modeling each collective activity class with an interaction matrix to measure connections between atomic activities.
• Formulating an interaction response (IR) model to create class-specific and distinct interaction patterns.
• Utilizing a multitask IR to jointly learn various interaction patterns and their relationships.

Main Results:

• The proposed model effectively captures discriminative low-rank representations of person-person interactions.
• Experimental results on two datasets show the model is comparable to state-of-the-art methods.
• Demonstrated that learning person-person interactions is critical for accurate collective activity recognition.

Conclusions:

• The novel learning-based approach significantly enhances collective activity recognition by focusing on interaction patterns.
• The method provides a robust way to model and exploit inter-person dynamics for group activity understanding.