Enhancing LGMD's Looming Selectivity for UAV With Spatial-Temporal Distributed Presynaptic Connections.

Area of Science:

• Computational Neuroscience
• Robotics
• Bio-inspired Engineering

Background:

• Collision detection is critical for unmanned aerial vehicles (UAVs), especially small ones with limited processing power.
• Flying insects, like locusts, exhibit efficient collision avoidance using motion-sensitive neurons, such as the Lobula giant movement detector (LGMD).
• Existing LGMD models struggle to differentiate genuine collision threats (looming) from background motion during agile UAV flights.

Purpose of the Study:

• To develop an improved LGMD-based collision detection model for UAVs.
• To enhance the model's ability to distinguish looming from complex background movements.
• To create a system suitable for implementation on resource-constrained embedded platforms for small and micro-UAVs.

Main Methods:

• Proposed a novel model inspired by locust synaptic morphology, incorporating distributed spatial-temporal synaptic interactions.
• Implemented locally distributed excitation to amplify motion detection at preferred velocities.
• Introduced radially extending temporal latency for inhibition to selectively suppress non-preferred motions and background clutter.

Main Results:

• The new model demonstrates significantly enhanced selectivity for looming stimuli over background movements.
• Spatial-temporal competition between excitation and inhibition effectively tunes the system to preferred image angular velocities.
• Experiments confirm improved performance in complex flying scenes representative of agile UAV maneuvers.

Conclusions:

• The proposed bio-inspired model offers superior looming selectivity for UAV collision detection.
• The model's design addresses limitations of existing LGMD approaches in dynamic environments.
• This system shows strong potential for practical application in embedded collision avoidance systems for small and micro-UAVs.