Stochastic and Evolutionary Looming Detection under Visual Noise

Neural models inspired by the locust's lobula giant movement detectors (LGMD), noted for their low power consumption and high computational efficiency, have significantly advanced visual collision detection from image streams. However, their performance often deteriorates in visually noisy environments. Biological studies indicate inherent randomness in synaptic transmission, suggesting that introducing probabilistic modeling could more accurately represent biological uncertainty and improve robustness against noise. A preliminary study recently demonstrated that incorporating a Bernoulli-distribution probability could enhance the LGMD model's robustness under noisy visual conditions. To further investigate which probability distribution optimally improves looming detection performance, this study proposed integrating a Gaussian-distribution probability into an LGMD neural network model with ON/OFF-contrast channels. The parameters of this model were searched through evolutionary computation across diverse day and night collision scenarios. Compared with the previous work, the method demonstrated superior robustness in both realistic and artificially noisy environments, achieving an 83% improvement regarding the distinct ratio, a metric to quantify sensitivity to noisy signals. An interesting finding through tests in generalized scenarios indicated that while the introduction of probability significantly enhances LGMD model's performance, the specific type of probability distribution is less critical. Moreover, this research explored variations in probability parameters across the ON/OFF-channels and suggested that stochastic signal processing not only effectively simulates uncertainty in neuronal transmission but also modulates signal propagation strength. This dual functionality balances neural processing and significantly enhances the robustness of looming detection in noisy visual conditions.