Research On Feature Extraction And Classification Of DVS Events Flow

Dynamic vision sensor(DVS)is a type of sensor that captures merely the moving objects by only sending out spatiotemporal information of the activated pixels,of which the brightness change exceeds the predefined threshold.DVS's output is in the form of events flow that contains multiple events instead of frames,this brings DVS an advantage that it is able to catch moving objects with much less storage space.In recent years,a growing number of applications,like monitoring and autopilot,have taken DVS as the visual input device,which urges the manufacturers to produce DVS of higher resolution,whereas how to implement classification of events flow with both high accuracy and instantaneity becomes an urgent problem.This thesis tries to solve the above problem by extracting low-dimensional feature and improving classification process.When visualizing the events flow,the image should display texture of the goals.Local binary pattern(LBP)is a low-dimensional specialized texture feature,based on which this thesis proposes a new feature called event-driven local binary pattern(ELBP)by merely computing the texture pattern for pixels where event happens.Due to the sparsity in spatial distribution of events flow,ELBP can help reduce computational cost.Experiments show this new method also helps improve the classification accuracy.The mainly adopted algorithms used for classifying events flow are frame-driven,however,frame-driven methods have some inherent drawbacks.To avoid these drawbacks,this thesis makes improvements on feature extraction method and classifying procedure respectively.In the part of feature extraction method,this thesis proposes a new algorithm called eventdriven ELBP.This algorithm regards a newly-arrived event as the trigger for computation and obtains the ELBP feature directly from limited time-space local region.Besides,for the consideration of computational efficiency,some adjustments are made in ELBP feature extraction.This method allows extracting features and acquiring classification results at any time,therefore it is able to get any number of features and thus keeps away the problem brought by imbalance of dataset.Its effectiveness is proved through experiments.On classifying procedure,this thesis proposes a strategy called majority voting.Events flow often represents the same goals in a period of time,therefore utilizing features from multiple moments to vote for the final answer,which is the main idea of majority voting,instead of relying on one single moment helps improve the accuracy.Experiment results verify its validity and show robustness of classification accuracies.Combing the event-driven ELBP and majority voting strategy,the full picture of eventdriven classification system for events flow is shown.This very system is able to get rid of the limitations existing in frame-driven ones,it is also the exact solution for classifying the events flow accurately in real time.