vSTMD: Visual Motion Detection for Extremely Tiny Target at Various Velocities.

Area of Science:

• Computer Vision
• Biomimetic Systems
• Robotics

Background:

• Detecting extremely tiny (ET-) targets is difficult for feature-based models due to their subtle visual cues.
• Insect-inspired Small Target Motion Detector (STMD) architectures show promise but are limited by narrow velocity ranges.
• Real-world scenarios require motion detection models that handle diverse and unstable target dynamics.

Purpose of the Study:

• To develop a learning-free motion detection model for ET-targets effective across a wide range of velocities.
• To enhance motion detection accuracy, robustness, and computational efficiency for ET-targets.
• To provide a real-time solution for ET-target motion detection in dynamic and complex environments.

Main Methods:

• Introduced vSTMD, a novel learning-free model incorporating cross-Inhibition Dynamic Potential (cIDP) and Collaborative Directional Gradient Calculation (CDGC).
• cIDP enables self-adaptive motion cue capture across a broad velocity spectrum.
• CDGC improves orienting accuracy and robustness while reducing computational load.

Main Results:

• vSTMD and vSTMD-F achieved relative F1 gains of 30% and 58% over state-of-the-art STMD methods on the RIST dataset.
• Demonstrated competitive orientation estimation performance against deep learning methods.
• vSTMD operates 60x faster than contemporary data-driven methods, highlighting its real-time applicability.

Conclusions:

• The proposed vSTMD model effectively addresses the velocity range limitations of previous STMD approaches for ET-target motion detection.
• Natural architectures like vSTMD offer a superior, faster alternative to data-driven methods for real-time ET-target tracking.
• vSTMD is highly suitable for real-time applications in dynamic scenarios and complex backgrounds due to its speed and accuracy.