Attention and Prediction-Guided Motion Detection for Low-Contrast Small Moving Targets

Area of Science:

• Robotics
• Computer Vision
• Neuroscience

Background:

• Autonomous robots struggle with detecting small targets in complex natural environments due to low visual contrast.
• Existing small target motion detection (STMD) models fail in natural settings where targets have minimal contrast.
• Insect visual systems exhibit remarkable efficiency in detecting small moving targets, even at low contrasts.

Purpose of the Study:

• To develop an advanced visual system for autonomous robots capable of detecting small, low-contrast moving targets in complex environments.
• To overcome the limitations of current STMD models that are contrast-dependent.
• To emulate the efficiency of insect visual systems for enhanced target detection.

Main Methods:

• Developed a recurrent visual system integrating three subsystems: attention, STMD-based neural network, and prediction.
• The attention module identifies potential targets and enhances their contrast.
• The STMD network processes enhanced images to detect motion, while the prediction module guides attention.
• The system processes information sequentially for targeted small target detection.

Main Results:

• The proposed attention-and-prediction-guided visual system significantly improves the detection of small, low-contrast moving targets.
• Experiments on synthetic and real-world datasets confirm the system's effectiveness and superiority over existing methods.
• The system successfully discriminates targets from complex backgrounds, addressing low-contrast challenges.

Conclusions:

• The developed visual system offers a robust solution for small target motion detection in challenging natural environments.
• The integration of attention, STMD, and prediction mechanisms provides a powerful approach inspired by biological vision.
• This research advances autonomous robot capabilities in complex visual perception tasks.