| Small target detection and tracking is one of the challenging research topics in computer vision,and the construction of appropriate computational models to detect and track small target objects efficiently and accurately is of great significance to the development of artificial intelligence.In recent years,the neural mechanism of visual perception in biological vision systems has been initially revealed,which provides new ideas to solve the difficult problems of visual perception in computer vision.At present,the neural structural properties of biological vision systems and response mechanisms have been initially applied to build new artificial vision systems which solve visual perception problems such as collision detection and motion pattern recognition of motion targets,but the detection of small target objects and motion tracking are still rarely reported.Therefore,it is of great scientific significance and potential engineering applications to deeply explore the neural structural properties of biological visual systems and response mechanisms,and to construct bionic artificial visual neural network models and algorithms for small target detection and tracking.Based on neurophysiological theories of locust vision,with the help of the neural structural properties of biological visual systems and response mechanisms of visual neurons,this thesis explores the corresponding artificial visual neural network models and algorithms that are applicable to the small target pedestrian detection and recognition and small target motion tracking in visual scenes,and the performance of computational models and algorithms.The main research works and results are as follows.1)For the problem of detecting small target pedestrian in visual scenes,an artificial visual neural network model and the corresponding algorithm are designed,based on the neural structural properties of locust visual systems,with the help of the mechanisms related to the decoding and encoding of episodic memory in human brains.The network consists of presynaptic and postsynaptic sub-neural networks.The former captures visual motion information elicited by small target pedestrians in visual fields;the latter extracts spike activity sequences that characterize the motion target and the small target pedestrian’s episodic memory,respectively,and compares them to detect small target pedestrian in visual fields.Theoretical analysis shows that the computational complexity of the proposed algorithm is determined by the resolution of the input video image frames.Systematic experiments display that the proposed model can effectively detect and recognize small target pedestrian in visual scenes.2)For the problem of small target motion tracking in visual scenes,an artificial visual neural network model and the corresponding algorithm are constructed,based on the neural structural properties of locust visual systems,with the aid of Drosophila ON-OFF channels for visual information processing and the predictive gain modulation mechanism in the dragonfly brains.The model extracts visual cues of the motion small target in visual fields that characterize the spatial position and match the direction of priori path,respectively,and integrates them to obtain spatial motion trajectory points depicting future moments for tracking small targets.Theoretical analysis shows that the computational complexity of the proposed algorithm is determined by the resolution of the input video image frames.Systematic experimental results claim that the proposed model can predict the motion trajectory of small targets in visual fields for motion tracking. |
