Reconstruction-Free Action Inference from Compressive Imagers.

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• Persistent surveillance generates massive video data, posing storage, communication, and computation challenges.
• Compressive cameras offer a solution but typically require video reconstruction for inference, which is computationally intensive and yields low-quality results at high compression rates.

Purpose of the Study:

• To develop reconstruction-free methods for action recognition from compressive camera measurements.
• To enable efficient action recognition at high compression ratios (100x and above).

Main Methods:

• Proposed spatio-temporal smashed filters, a compressive domain adaptation of pixel-domain matched filters.
• Developed methods to extract features directly from compressive sensing (CS) measurements, bypassing traditional reconstruction.

Main Results:

• Achieved action recognition rates comparable to the oracle method (uncompressed setup) even at high compression ratios.
• Demonstrated the effectiveness of reconstruction-free inference for compressive cameras.

Conclusions:

• Reconstruction-free methods are viable and efficient for action recognition with compressive cameras.
• Spatio-temporal smashed filters effectively capture necessary features directly from compressive measurements.