Fly Visual Brain Response Mechanism-Inspired Neural Network Construction And Related Motion Collision Avoidance System Application

Collision avoidance is an important puzzle concerned in the fields of science and technology such as robotics,unmanned aerial vehicles,and intelligent transportation.It involves how to timely detect collision hazards and accurately transmit danger signals before moving objects encounter collision and effectively avoid accidents as well,which belongs to the research branch of computational intelligence and hybrid systems.Based on computational geometry,laser,ultrasound,radar,etc.,traditional computing models on collision avoidance are difficult in handling hazard avoidance when one or more objects appear in the whole visual region,due to low object detection accuracy,high cost,and limited application scope.This greatly restricts their wide applications.Accordingly,the main concern in the dissertation is how to explore new-type,rapid,and effective artificial visual brain computing models to quickly and effectively implement collision detection and motivate the moving object(s)in the visual scene to avoid potential hazards.As a model organism with a relatively simple visual brain structure,flies have been comprehensively studied in neurophysiology.Their response properties can provide us with a new way to study neural computing models for collision detection,hazard avoidance,etc.However,it is a still brand-new research topic how the information-processing mechanism and response characteristics of the fly visual brain system can be used to construct collision detection models,collision avoidance schemes,omnidirectional collision detection neural networks,and intelligent collision avoidance systems.Therefore,related to fly neurophysiologic findings,this dissertation solves the problem of omnidirectional collision detection and collision avoidance and achieves an intelligent collision avoidance system.To that end,a series of works have been carried out about neural computing model construction and design,theoretical analysis,performance test,and Simulink simulation,in order to execute collision detection and collision avoidance in forward wide or whole regions.The acquired achievements not only can provide new schemes to solve the problems of robot navigation,UAV,and traffic safety at intersection in the whole region,but also have an important contribution to the rapid development of artificial intelligence and computer vision.The main work and achievements acquired are summarized below:A.Based on the inspiration of the neural structure of fly visual information processing and the visual response characteristics of motion-sensitive neurons,a feedforward visual neural network with presynaptic and postsynaptic neural subnetworks is developed to the changes of visual motion in the forward wide region.After that,such a visual neural network,together with a forward collision detection model,is used to construct a forward collision detection system capable of performing collision detection in the half-plane region,particularly edge motion detection.In the designs of the model and system,two parabolic curves are used to divide the local motion direction detection layer into the left,forward and right visual sub-regions,by means of the perceptual properties of motion-sensitive neurons which only respond to their preferential visual regions.Based such a regional division,the presynaptic neural network consists of three visual neural subnetworks in order to extract the changes of visual motion in the left,forward,and right visual regions and output the related neurons’ activities to measure the state of visual motion in the forward wide region scene.The postsynaptic neural network,which includes the visual neural subnetworks’ output neurons and a spike detection neuron,outputs its on-line activities and spike transmission signals in order to measure the changes of visual motion in the visual scene.Accordingly,a forward collision detection system,which can detect the state of danger occurrence in the forward wide region scene and transmit collision early warning signals on line,is formed by the feedforward visual neural network and a collision detection model.Theoretical analyses show that the computational complexity of the collision detection system is mainly determined by the resolution of each input image frame.Comparative experiments and Simulink simulation have demonstrated that the acquired visual neural network can characterize the visual perceptual properties of motion-sensitive neurons,and that the collision detection system can effectively detect collision hazards generated by moving objects and transmit collision early warning signals.B.Based on the preferential response characteristics of horizontal and vertical motionsensitive neurons and spike transmission ones in the fly visual system,a visual neural network is formed of the presynaptic and postsynaptic neural networks.The presynaptic neural network with four sub-neural networks is developed to detect the changes of visual motion appeared in the left,right,forward,backward subregions of the whole visual region,after extending the feedforward visual neural network in section A above.Therein,two diagonal lines are used to divide the local motion direction detection layer into four subregions,and meanwhile a symmetrical lateral inhibition mechanism is used to compute the activities of nodes in the local motion direction detection layer.The postsynaptic neural network,which involves in the four output neurons of the presynaptic neural network and a spike detection neuron,is developed to detect the changes of visual motion and perform omnidirectional radial collision detection in the whole region.Related to the visual neural network and a collision detection model,the omnidirectional collision detection system,which consists of the visual neural network and a collision detection model with the three neurons on feedforward inhibition,spike detection,and danger detection,is developed to perform omnidirectional collision detection.Therein,the feedforward inhibition neuron is used to detect whether the visual scene causes the phenomena of motion vibration and bright light interference.Theoretical analyses show that the computational complexity of the system is determined by the resolution of the input image and the lateral inhibition radius between adjacent neural nodes.Comparative experiments and Simulink simulation have illustrated that not only the acquired visual neural network can well characterize the specific response properties of the visual neurons,but also the system can effectively perform collision detection and timely transmit collision early warning signals with a high success rate of collision detection.C.For the problem of collision detection of equatorial and omnidirectionally radial motion in the whole region,a visual brain neural network with presynaptic and postsynaptic neural networks is proposed to detect the changes of angle and height of the main moving object in the whole visual region.On the basis of simplifying the visual neural network in section B,the presynaptic neural network with five neural subnetworks is developed to detect the changes of radial and equatorial motion in the panoramic scene,in which the response characteristics of the equatorial motion-sensitive neuron are adopted to construct a computational model capable of detecting the pattern of equatorial motion.The postsynaptic neural network,which includes the five output neurons of the presynaptic neural network,two neurons of angle and height detection,and the structures of the protocerebral bridge and fan-shaped body,is responsible for detecting the changes of the main moving object in the panoramic scene.Since the angle and height sensitive neurons in the fly brain nervous system have specific response characteristics to the visual angle and height of the main object,they are used to construct computational models used to estimate the changes of the moving object.Subsequently,a collision detection system,which involves in the visual brain neural network and a collision detection model,is developed to detect the changes of equatorial and omnidirectionally radial motion and transmit collision early warning signals on line.In the designs of the model and system,the rectangular region of the local motion direction detection layer is divided into five subregions,i.e.forward,backward,left,and right subregions and equatorial detection subregion,by means of two symmetric parabolic curves and an elliptical curve.Each subregion,together with one of the above five motion-sensitive neurons and the three visual neural information processing layers,constitutes a presynaptic neural subnetwork.Subsequently,based on the symmetrical inhibition and cross transmission mechanisms in the protocerebral bridge and the response characteristics of angle and height sensitive neurons in the fan-shaped body,the postsynaptic neural network is designed to measure the changes of the visual height and angle of moving objects in the whole region.On the other hand,based on the acquired visual brain neural network and three neurons on feedforward inhibition,synthetic spike inhibition,and danger transmission,a collision detection system is constructed to transmit danger signals.Theoretical analyses have validated that the computational complexity of this system is determined by the size of the resolution of the input image frame and the size of the radius of symmetric lateral inhibition between the neural nodes.Comparative experiments and Simulink simulation have illustrated that the output of the acquired visual neural network can well characterize the response properties of its output neurons,and meanwhile the system can timely and effectively transmit collision early warning signals.D.Aiming at the problem of omnidirectional collision avoidance of moving objects,a feedforward fly visual brain collision avoidance neural network with presynaptic and postsynaptic neural networks,which bases on the design of the neural network in section C,is constructed to transmit collision avoidance instructions and output excitatory activities which characterize the changes of omnidirectional motion in the panoramic scene.Herein,some prominent response characteristics,decision-making neural circuits,and circular neurons in the brain nervous system are borrowed to construct neurocomputational models which appear in the postsynaptic neural network.On the other hand,an intelligent collision avoidance system,which consists of the abovementioned collision avoidance neural network and a collision avoidance model,is designed to perform collision detection and collision avoidance in the whole visual region.Herein,relying upon the response characteristics of the collision avoidance neuron(s)and the information processing mechanism of the neural pathway in the aforementioned brain nervous system,the collision avoidance model is developed to perform collision avoidance when a danger occurs.In the designs of the model and system,the response characteristics of motion-sensitive neurons are used to divide the rectangular region of the local motion direction detection layer into four overlapping perception regions with each other in terms of four parabolic curves.Such four regions,together with three visual neural information processing layers and four motion-sensitive neurons,constitute a presynaptic neural network with four neural subnetworks in order to perceive the changes of visual motion in the whole visual region.Related to the neural information transmission and processing mechanisms of the protocerebral bridge,fan-shaped body,and ellipsoid body in the central complex,the postsynaptic neural network is developed to detect the changes of the main moving object in the whole region.After that,associated to the response characteristics of dopaminergic neurons and the mechanisms in the mushroom body compartment,decision-making neural pathways,a collision avoidance model is designed to deliver collision avoidance commands and compute collision avoidance risk values.Herein,the characteristics of the effective refractory period of neural cells,the logic mechanism of the XOR logic operation neuron and the control mechanism of the thoracic locomotor control center are used to achieve the functions of the model.Theoretical analysis shows that the computational complexity of the collision avoidance system is determined by the resolution of the input image and the symmetric lateral inhibition radius between the neural nodes.Comparative experiments and Simulink simulation have validated that the visual brain collision avoidance neural network can exhibit some functional characteristics of collision perception and collision avoidance command transmission;particularly,the intelligent collision avoidance system can timely discover the related collision region,transmit collision avoidance instructions,and implement collision avoidance.