Research And Application On Motion Pattern Recognition Based On Locust Vision Neural Systems

Motion pattern recognition is a challenging research branch in the field of computer vision.It extracts the behavior characteristics of motion objects in the field of view,by designing characteristic recognition computational models which originate from some theories and methodologies in computer vision,computation geometry,neural network,visual neurophysiology,etc.To this end,an extremely important topic is how to develop efficient computational models and probe into appropriate algorithms for motion pattern recognition,in order to promote the development of artificial intelligence.For a moving object,traditional computer vision techniques can catch some of its characteristics but are difficult in recognizing its motion patterns.However,studies on visual systems with efficient and robust motion pattern recognition have made a dent in neurophysiology,which can provide inspirations from the angle of computer vision to build computational models for such motion perception tasks as motion pattern recognition,target tracking,collision detection,etc.Generally,the diversity of motion patterns in the real world can be decomposed into translational,radial,rotational,depth-rotation and spiral motion.Interestingly,in biological vision systems,scientists have found that specific types of visual motion perception neurons have specific preferences for motion patterns.The response mechanisms of visual neurons in animal's brain have been applied to developing artificial vision systems suitable for translational and radial motion pattern recognition.However,little works have been done in the past to create computational models for other motion patterns.It is a still urgent topic that researchers probe into bio-plausible computational models to recognize the motion patterns of rotational,depth rotation and spiral motions.Therefore,it is of importantly scientific and practical significance to develop artificial visual neural networks that can not only simulate the visual response characteristics of animals but also solve motion pattern recognition problems,based on the hierarchical frameworks and response mechanisms of visual motion perception neurons in biological vision systems.As associated to the neurophysiological theories of vision systems of locust and macaque,as well as the biological theories of visual perception and motion pattern recognition,this dissertation investigates artificial visual neural networks and their corresponding algorithms used for the recognition of rotational,depth rotation and spiral motion.Meanwhile,such neural networks and algorithms are also sufficiently analyzed with the aspects of computational complexity,comparative analysis,parameter sensitivity,algorithmic applications and so forth.The main works and achievements can be summarized below.A.In order to recognize rotational motion presented in visual scenes,an artificial visual neural network and the related algorithm are developed,based on the structural characteristics of the locust vision neural system and the visual response mechanisms of rotation sensitive neurons in macaque's cerebral cortex.This neural network consists of presynaptic and postsynaptic neural networks.The presynaptic network includes sixteen directional selective neural networks with similar internal structures but different functions.To perceive the continuous change of motion direction,these directional selective neural networks are arranged in a cyclic structure similar to the arrangement of direction columns in the mammalian's cerebral cortex.The postsynaptic network,however,gathers the outputs of such direction selective neural networks by means of direction column,while synthesizing such outputs to form the output membrane potentials which can represent the behavior characteristics of rotational motion.In the design of the algorithm,different asymmetric lateral inhibitory mechanisms are used to develop computational models for the perception of different direction cues.Also,the visual response mechanisms of rotation sensitive neurons and the mechanism of the Reichardt motion detector are utilized to create computational model for rotational motion pattern recognition.Theoretical analyses indicate that the computational complexity of the proposed algorithm is determined by the resolution of video input images.The comparative experiments,based on rotational and non-rotational motion video sequences in different scenes show that the proposed neural network can not only efficiently reveal the specific visual perception characteristics of rotation sensitive neurons,but also effectively recognize rotational motion patterns.B.For the problem of depth rotation recognition presented in visual scenes,an artificial visual neural network,which is composed of presynaptic and postsynaptic networks,is developed,based on the visual characteristics of the locust vision neural system and the visual response mechanisms of depth rotation sensitive neurons in macaque's cerebral cortex.In the neural network,the presynaptic network comprises of a depth perception neural network and eight directional selective neural networks for capturing visual information along different motion directions;the postsynaptic network collects the outputs of the presynaptic network by means of direction column,while synthesizing them to perceive the spatio-temporal energy change of depth rotation.In the design of the algorithm of the visual neural network,one computational model,which catches different translational direction cues,is established by utilizing the visual characteristics of directional selective neurons;one model used for collecting looming/receding motion cues is designed based on a symmetric lateral inhibitory mechanism.Additionally,in order to detect the spatio-temporal energy change of depth rotation,one model,based on the visual response mechanism of depth rotation sensitive neurons and the neural structure characteristics of direction column is developed to fuse the captured visual motion cues.Theoretical analysis shows that the computational complexity of the proposed algorithm is determined by the resolution of video input images and the inhibition radius of directional selective neurons.Numerical experiments,based on video sequences in different scenes validate that not only the proposed neural network can effectively perceive the spatio-temporal energy change of depth rotation,but also its output excitation curve is a quasi-sinusoidal one.This is compatible with the hypothesis of depth rotation perception from projective geometry studies.C.For the problem of spiral motion recognition presented in visual scenes,an artificial visual neural network and its related algorithm are investigated,based on the structural characteristics of the locust vision neural system and the visual response mechanisms of spiral motion selective neurons in macaque's cerebral cortex.One such network,inspired by two stages of biological visual information processing includes two subnetworks —presynaptic and postsynaptic networks.The former includes one feed forward neural network for extracting radial motion cues and sixteen directional neural networks for capturing translational direction cues.By means of the structural characteristic of direction column,the latter collects visual motion cues from the presynaptic network and synthesizes them to recognize the pattern of spiral motion.In the design of the algorithm,one computational model,based on symmetric lateral inhibitory mechanism is designed to perceive the radial motion cues;one model is developed to perceived translational direction cues,by means of one asymmetric lateral inhibitory mechanism;one model is constructed to extract different rotational and radial motion cues,relying upon the neural characteristics of direction column.Finally,one model,based on the visual response mechanism of spiral motion selective neurons is designed to recognize spiral motion pattern.Theoretical analysis manifests that the computational complexity of the proposed algorithm is determined by the resolution of the video input image and the inhibition radius of directional selective neurons.Depending on spiral and non-spiral motion video sequences in different scenes,numerically comparative experiments show that the proposed neural network can not only explain the visual response characteristics of spiral motion selective neurons,but also effectively recognize spiral motion patterns.D.In order to recognize the abnormal activities of crowd escape behavior presented in video monitoring systems,an improved LGMD neural network,which can depict crowd behavior characteristics in the field of view is developed,based on the visual response mechanisms of the lobula giant movement detector neurons in the locust lobula.After that,a unique adaptive threshold scheme is designed to regulate the discharge excitation of the neural network for the recognition of the process of the crowd escape behavior.Theoretical analysis indicates that the computational complexity of the algorithm derived from one such neural network is determined by the resolution of video input images.Numerical experiments have validated that the neural network,based on the UMN data set of crowd scene videos is effective and reliable in predicting crowd escape behavior.