| With the vigorous development of large-scale models such as GPT(Generative Pre-trained Transformer)and the widespread adoption of AIGC(Artificial Intelligent Generated Content)technology,deep learning-based artificial intelligence research demonstrates powerful potential applications in many fields.However,some drawbacks are beginning to emerge.On the one hand,the large number of determinations of weighting parameters in deep networks is inseparable from massive training data.On the other hand,the massive computing requirements of large models are increasingly highlighting energy consumption issues.The dilemmas faced by deep learning technology bound the further development of artificial intelligence.Therefore,exploring a new research path for artificial intelligence is imperative.In nature,small insects exhibit amazingly intelligent behaviors in basic life activities such as foraging,mating,and avoiding enemies,based solely on their limited micro-brain network architecture with extremely weak neural computational energy consumption.Based on this,in recent years,the analysis of the precise neural computation behind the micro-brains of small insects has gradually become a novel research paradigm for constructing a new generation of artificial intelligence.In the field of insect micro-brain intelligence research,visual intelligence plays a crucial role.Neuroscience research shows that flying insects such as locusts exhibit outstanding intelligent collision avoidance behavior in swarm migration due to the sensitive and robust responses by LGMD(Lobula Giant Movement Detector)neurons in the lobula to looming visual stimuli.Therefore,exploring and modelling computational mechanisms of LGMD neurons opens up a whole new avenue for building “low-computing-power,low-energy,non-learning-based”collision-detection vision systems.Although researchers have established several neural network models based on LGMD neurons over the past few decades,these models are all based on monocular vision.It is well known that binocular vision is the inherent visual structure of many organisms of nature.Indeed,from a neural anatomical perspective,the central complex of the locust brain has symmetrical visual lobe structures on either side connecting the left and right compound eyes,and physiological and behavioral experiments also confirm the fundamental importance of binocular visual information for the collision avoidance ability of locusts.However,how binocular visual information is fused,particularly how it manifests its functional role in LGMD artificial neural network models,has not been investigated.In this thesis,we investigate this issue in depth,and the main findings are as follows:(1)A basic binocular vision-based LGMD collision detection model is proposed,which solves the novel problem of designing a binocular information fusion mechanism within the classical framework of existing models.Compared to monocular vision,the fundamental advantage of binocular vision lies in the perception of depth-distance.This article effectively extracts depth-distance information of moving objects using a binocular information fusion mechanism,improves the model’s ability to discriminate basic motion patterns,and shapes clear response selectivity.Secondly,this article introduces a dynamic adaptive depth-distance warning mechanism to ensure that the model can always generate collision warning responses within a relatively reasonable time.A series of numerical experimental results illustrate that the proposed model has a more potent comprehensive performance.(2)A binocular visual-based LGMD collision detection model with approaching azimuth recognition is proposed,addressing the inherent problem of the existing models’ inability to differentiate the inherent approaching azimuths with different collision risk levels and improving the response selectivity and rationality of the models.The approaching azimuth,one of the essential features of motion information,still struggles to be characterized by the existing models,resulting in their inability to respond reasonably to different approaching visual stimuli.Therefore,the research approach is from specific to general in this paper.First,for the typical case of the frontal approaching directly,the approach azimuth problem is transformed into an approach angle problem,therefore designing a binocular LGMD collision detection model with approach angle estimation.Second,for the more general case,this article combines spatial projection transformation to propose a binocular LGMD collision detection model with approaching azimuth and spatial position estimation,which deepens the existing LGMD model’s intelligent understanding of approaching patterns.A series of numerical experimental results demonstrate the effectiveness and superiority of the proposed model.(3)A binocular visual-based LGMD collision detection model with single peak response morphology is proposed,cracking the activation threshold dependency problem of the existing models and improve the model’s robustness to visual factors.The existing models are sensitive to activation threshold influences and show unstable performance under variable visual factors such as contrast and noise.Therefore,this paper develops a -type activator,inspired by the function,to shape the response results into a single-peak morphology and ensure that the peak appears before the collision moment.Numerical experimental results show that the peak-based collision warning response mechanism is more robust to variable visual factors.Simulation tests on the WEBOTS platform also confirm that the miniature visual robot equipped with the proposed model exhibits more stable collision avoidance performance in different visual environments.The neural computation method proposed in this paper opens a new direction for designing intelligent units for miniature visual robots.(4)A binocular visual-based LGMD collision detection model against extreme noise is proposed,solving the problem of inadequate performance of existing models and the above binocular LGMD models under extreme noise conditions.This paper introduces pooling units between neural network layers and implements effective extraction of collision signals under extreme noise conditions through a competition mechanism between maximum and average pooling.Numerical experimental results also demonstrate the advantage of the pooling competition mechanism over traditional filtering methods in the LGMD visual neural network model.In addition,the introduction of pooling units in the LGMD model considerably reduces the computational complexity of the model,thus providing a new direction for exploring lightweight and efficient noise-resistant LGMD visual collision detection models. |
