Denoising Methods For Address Event Stream Data Of Dynamic Vision Sensor

At present,the traditional image sensor has been widely used,which greatly facilitates People's Daily life.However,This kind of sensor usually records the shooting scene with fixed frame rate,which has the disadvantages of long response time,high time delay and lots of information redundancy.Dynamic Vision Sensor(DVS)is a new Sensor driven by bionic asynchrony.Each pixel of the Sensor independently monitors the change of external light intensity and generates events when the change in light intensity(increase or decrease)exceeds the set threshold.Because of its short response time,DVS can monitor the high-speed motion information captured by expensive high-speed cameras in traditional methods(tens of thousands of frames per second),and its data volume is reduced by thousands of times,which greatly reduces the cost of subsequent signal processing.The technology has a wide range of applications in aerospace,automotive autonomous driving,consumer electronics,industrial vision and safety.The high sensitivity of DVS gives it a natural advantage in capturing fast-moving objects,but it also provides low noise robustness.Due to the sensor's own components,circuit design,external interference and other reasons,the address event stream data output by DVS often contains a large number of noise events.Different from the traditional two-dimensional frame image data format,DVS outputs data in the form of Address Event Representation(AER),which requires a new address event stream data denoising method.Based on the imaging principle of DVS and the characteristics of event stream data,this paper proposes two kinds of address event stream data denoising methods for dynamic vision sensors.The first one is the address event stream data denoising method based on temporal spatial filtering.This method makes full use of the special time information of the event,and combines the randomness and isolation of the noise event to construct a three-dimensional spatio-temporal filter for denoising.The denoising method directly removes the noise event from the address event stream data to ensure the information integrity of the address event stream data.The proposed algorithm is simple in principle,fast in speed,and can effectively remove noise events.However,the denoising effect is not ideal when dealing with complex scenes with high noise.The second is the address event stream data denoising method based on the probabilistic undirected graph model.This method extends the regularity of object motion to the spatio-temporal correlation between effective events,and combines the priori knowledge such as the randomness of noise events to construct a probabilistic undirected graph model,thus transforming the denoising problem into a probabilistic maximization problem.Then the probabilistic undirected graph model is decomposed into the product form of the energy function on the maximum cliques,and the probability maximization problem is transformed into the energy minimization problem.Finally,the Iterated Conditional Modes(ICM)optimization algorithm is used to optimize the model,so as to realize the denoising of address event stream data.At the same time,adaptive dynamic weight parameters of space domain and time domain are designed to improve the model.This denoising method has good robustness and can deal with the denoising problem of address event stream data in multiple complex scenarios.