A spatiotemporal energy model based on spiking neurons for human motion perception

Area of Science:

• Computational Neuroscience
• Computer Vision
• Biophysics

Background:

• The human visual system excels at processing motion, a capability crucial for tasks like navigation and action recognition.
• Existing computational models often simplify neuronal dynamics, potentially limiting their biological realism and performance.

Purpose of the Study:

• To develop a bio-inspired feedforward spiking neural network model for human motion perception.
• To replicate the direction selectivity mechanisms of simple and complex cells in the primary visual cortex.
• To evaluate the model's performance in motion detection and segmentation tasks.

Main Methods:

• Utilized Hodgkin-Huxley neurons to model spiking neuron dynamics.
• Modeled receptive fields of simple cells using Gabor energy filters.
• Constructed complex cell receptive fields by integrating simple cell responses within an energy model.
• Generated motion maps by integrating directional motion information from complex cell outputs.

Main Results:

• The spiking network successfully replicated the directional selectivity of the visual cortex for time-varying images.
• The model demonstrated effective motion energy extraction from diverse video sequences, comparable to human visual perception.
• Achieved competitive performance in motion segmentation tasks against state-of-the-art models.

Conclusions:

• The proposed bio-inspired spiking network offers a biologically plausible and effective approach to human motion perception.
• The generated motion maps are suitable for downstream applications such as action recognition.
• This model advances the understanding and computational modeling of visual motion processing.