Research On Object Detection And Recognition Algorithm Based On Dynamic Vision Sensor

With the development of science and technology,the accuracy and speed of object detection have gradually improved,but its effect in high-speed motion scenes,bright or dark scenes is not satisfactory.Event cameras,which have emerged in recent years,can perceive changes in light intensity and generate events.It can solve the problem of insufficient light conditions and the inability to capture high-speed motion.Therefore,this thesis uses Dynamic Vision Sensors to propose two event-based object detection algorithms.The algorithm has made innovations and improvements in event processing,feature extraction,and feature classification.First of all,this thesis introduces the types,working principles and advantages of the event camera,and establishes the mathematical model of the Dynamic Vision Sensor.Through the analysis of the events,we proposed methods from two perspectives,discrete and continuous.The event processing models are named the Integral model and the Leaky Surface model.Secondly,this thesis proposes two event-based object detection algorithms,namely event-based improved HOG+SVM object detection algorithm and event-based deep learning object detection algorithm Event-YOLO.The improved HOG+SVM object detection algorithm mainly converts the event into a grayscale image through the event processing model,and uses the improved HOG feature for feature extraction,which strengthens the overall feature extraction and reduces the attention of local detail features.Finally,through improved SVM,multiple objects are detected and classified.EventYOLO is a model which adds Leaky Surface model on the basis of YOLO network,realizing the processing of events.In addition,the event-based convolutional layer,the maximum pooling layer and the fully connected layer are designed and combined to realize the event-based target detection algorithm.Finally,this thesis uses MINST-DVS,POKER-DVS,CIFAR10-DVS and combined MINST-DVS event data set.Through the modification and adjustment of the data set,this thesis realizes the function of single object detection,multi object detection,fast motion scenes and actual life background.In the MINST-DVS and combined MINST-DVS data sets,the detection accuracy rate is more than 90%,which verifies the feasibility of the Dynamic Vision Sensor in the field of object detection.