Research On Crowd Gathering And Pedestrian Retrograde Behavior Detections Based On Locust Vision Neural Systems

Detecting crowd gathering and pedestrian retrograde behaviors in moving crowd is a key technology for video surveillance in public places,and it is also an important topic in computer vision.Constructing appropriate computational models and exploring suitable algorithms to achieve efficiently and accurately automatic detection of such abnormal events in crowd mentioned above are important for future artificial intelligence.However,due to the strong randomness and inconspicuous behavior features in crowd activities,traditional computer vision technologies are hard to recognize behaviors of crowd activities and weak adaptability to environments.In recent years,bionic computational models based on neural mechanisms of biological vision perception have been initially applied to visual perception and detection tasks,such as target detection and motion pattern recognition,but the study of abnormal behavior patterns,including the gathering and retrograding activities in moving crowd,is still a challenging issue to be observed.Therefore,it is of great academic significance and potential engineering application to build artificial visual neural network models that simulate structural characteristics and neural response mechanisms of biological vision systems for solving the problems of detecting abnormal behaviors in crowd activities.Based on visual neurophysiological theories of locusts,as well as neural structure characteristics of biological visual systems and response mechanisms of visual neurons,this thesis investigates the construction of artificial visual neural networks and algorithms that are applicable to crowd gathering behavior detection and pedestrian retrograde behavior perception in public places,and the performance of computational models and algorithms for engineering applications.The main works and achievements are as follows:1)For the problem of detecting sudden crowd gathering behaviors in public places,an improved artificial visual neural network and the corresponding algorithm are developed,based on the neural structures of locust vision systems and the visual response properties of lobula giant movement detector neurons in locust optic lobe.In the proposed neural network,visual signals are firstly collected in the field of view.With the help of visual information processing mechanisms of the mammalian retina for integrating the motion information of each pixel and obtaining convergent cues of the spatio-temporal energy change,a spiking threshold mechanism is developed to tune the global spatio-temporal spike response output for the perception of sudden gathering behaviors in crowd activities.Theoretical analysis displays that the computational complexity of the proposed algorithm is determined by the resolution of input video frames.Numerical experimental results of crowd activity video sequences in different real scenes indicate that the neural network can efficiently simulate the visual response properties of hazardous perception in lobula giant movement detector neurons,and also can effectively detect the sudden gathering behaviors of crowd activities in the field of view.2)For the problem of detecting pedestrian retrograde behaviors in crowd activities,an artificial visual neural network and the corresponding algorithm are designed,based on the structure characteristics of locust vision systems and the visual response properties of direction contrast sensitive neurons in the intermediate/deep layers of barn owls’ optic tectum.The proposed neural network consists of presynaptic and postsynaptic sub-neural networks.The former is composed of four directional selective neural networks,which extracts motion direction cues in the basic translation directions by the combining neural mechanisms of symmetric lateral inhibition,asymmetric lateral inhibition and directional column.The latter converts motion direction cues into directional contrast operations,and detects the pedestrian retrograde behaviors in the field of view through the spiking threshold mechanism.Theoretical analysis shows that the resolution of input video frames determines the computational complexity of the proposed algorithm.Numerical experimental results on the video sequences of moving pedestrian behaviors in different real scenes demonstrate that the network can not only effectively reveal the visual response properties of direction contrast sensitive neurons,but also can effectively perceive the retrograde pedestrians in moving crowd.