Retrieving Object Motions From Coded Shutter Snapshot in Dark Environment.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Traditional video object detection struggles with low image quality and motion blur, especially in dark environments.
• Limited sensor sensitivity in extreme darkness forces a trade-off between signal-to-noise ratio and motion blur compensation, degrading performance.
• Existing methods require high imaging quality and clear motion, limiting their application in challenging low-light conditions.

Purpose of the Study:

• To develop a robust video object detection method for extremely dark scenarios.
• To address the limitations of current detectors in low illumination and motion blur.
• To enable accurate object detection and trajectory retrieval from single, temporally multiplexed snapshots.

Main Methods:

• Temporally multiplexing a frame sequence into a single snapshot using a programmable shutter for encoded capture.
• Designing an end-to-end deep network, DEtection from Coded sNapshot (DECENT), for decoding and retrieving object trajectories from the coded snapshot.
• Generating quasi-real training data by incorporating physically-driven noise into a temporally coded imaging model.

Main Results:

• The DECENT network successfully retrieves sequential bounding boxes from coded blurry measurements of dynamic scenes.
• The approach demonstrates high accuracy and generalization ability on real dark videos, overcoming training data scarcity.
• Experimental validation confirms effectiveness under low illumination vision.

Conclusions:

• The proposed method provides a feasible solution for object detection and trajectory retrieval in challenging low-light and night surveillance scenarios.
• The approach offers advantages such as low bandwidth, low cost, and a compact setup.
• This technique significantly improves video object detection performance under extreme darkness and motion blur.