Research On Event-based Object Location And Recognition Model

The neuromorphic vision sensor is a bio-inspired sensor that simulates the biological retina.This new vision device is composed of a photoreceptor array.Each photoreceptor unit generates events asynchronously to encode the light intensity changes of the scene into the spatio-temporal event stream without frames.This kind of vision sensor has the advantages of low redundancy,low time delay,high time resolution,high dynamic range and so on.Because event stream and traditional image stream are essentially different in data format,researching and constructing a visual processing method suitable for event stream data is a major research topic of neuromorphic vision.Both the spiking neural network and the neuromorphic vision sensor simulate biological impulses,and they have the same information processing methods,so the events can be directly used as the input of the spiking neural network.The temporal coding method of the spiking neural network greatly retains the accurate time information of the events under the high time resolution.Researching a spiking neural network model for processing events recognition tasks is of great significance for promoting the development of neuromorphic vision.However,the existing event-based object recognition methods are not perfect,and many methods lead to the loss of accurate time information.In addition,these models are generally only effective on stable data sets with a fixed trajectory in a small range,and do not have the robustness of object recognition with a random trajectory.The movement is a necessary condition to trigger an event,but the real scene rarely meet trajectory completely fixed.The robustness of the event-based recognition model to object with a random trajectory is necessary for the real-world application.This paper constructs an event-based spiking neural network object location and recognition model.The main contents of this paper are as follows:1.This paper proposed a target feature extraction method based on the scale fusion of multi-spike events,and use the tempotron classification algorithm with population coding to construct an event-based object recognition spiking neural network.The model has achieved better classification effects than several current typical methods on multiple data sets.2.This paper proposed an activated connected domain algorithm for object location.By dynamically updating the active pixel set,the bounding rectangle of the active domain is obtained as the object location.The event-driven object location algorithm avoids the additional delay of most object location algorithms based on events.3.This paper proposed the integrated design of the object location and the recognition process,which realizes the robust object location process without an additional denoising module.An event-based target location and recognition model is constructed to realize robust recognition of random moving targets.