Learning and Inferring "Dark Matter" and Predicting Human Intents and Trajectories in Videos.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Human Behavior Analysis

Background:

• Objects in public spaces are often visually indistinct or occluded, hindering direct detection.
• Human movement patterns in public areas are influenced by unobserved functional objects, termed 'dark matter'.
• Existing methods struggle to infer object presence and human intent from trajectory data alone.

Purpose of the Study:

• To develop an unsupervised method for localizing functional objects ('dark matter') in surveillance videos.
• To predict human intents and future trajectories based on their interactions with these unobserved objects.
• To demonstrate the capability of inferring object influence from human movement patterns.

Main Methods:

• Modeling individuals as intelligent agents influenced by attractive/repulsive 'fields' generated by 'dark matter' objects.
• Utilizing Agent-based Lagrangian Mechanics to derive agent trajectories, allowing for dynamic intent changes.
• Developing an unsupervised learning framework for simultaneous object localization and trajectory prediction.

Main Results:

• Successfully localized and identified distinct types of 'dark matter' objects in public spaces.
• Achieved high accuracy in predicting human intents and future trajectories.
• Demonstrated the effectiveness of inferring object influence indirectly through observed human behavior.

Conclusions:

• The proposed method effectively addresses the challenge of detecting visually ambiguous objects by analyzing their impact on human trajectories.
• This approach offers a novel way to understand and predict human behavior in complex public environments.
• The findings open new avenues for surveillance analytics and human-object interaction studies.