The Design Of Events Feature Extraction Methods For The AER Vision Sensor System

Traditional vision sensors are based on the sampling mode of frame.As the development of the vision system applications,traditional vision sensors are limited in large data transmission,small frame frequency and low dynamic range.Therefore,AER(Address Event Representation)vision sensor based bionic visual perception mode has become a hot area of research.AER asynchronous event-driven vision sensor has the advantages of high-speed,low delay,and low redundancy.It is appropriate for high-speed objects tracking and target identification.In this paper,three events feature extraction methods based on AER vision system have been proposed,which could obtain the shape and texture features from the low-redundant events data in real time.Firstly,this paper briefly introduces the operating principles and the basic structures of the array AER vision sensor and the linear TAE vision sensor.Then,several shortcomings of AER vision system are shown.Aiming at the target shape feature extraction,a high-speed target binarization method for the linear TAE vision sensor is proposed in this paper.In the preprocessing phase,the events data are processed by denoising,thinning and edge connection methods sequentially to obtain the clear contour event-pair.Then the object regions are confirmed by an event-pair matching method.Meanwhile,a label-equivalence connected-component labeling algorithm based on AER encoding is designed.This algorithm only requires to label the AER events rather than the whole pixels in the image.At last,an AER convolution algorithm with 16 kinds of Gabor template to extract the texture feature of AER events is proposed.Through the experimental analysis and comparison with the traditional algorithms,simulation results show that the proposed binarization algorithm can deal with the complex illumination conditions such as non-uniform illumination and low contrast with high accuracy and efficiency.The runtime is about 2s~4s for the 512×512 object images.And the proposed AER connected-component labeling algorithm has a 1.5~8 times higher speed than traditional label-equivalence labeling algorithms.Furthermore,the AER convolution algorithm designed in this paper also could effectively extract the events texture features under the different directions and scales.To sum up,those three algorithms could effectively accomplish the feature extraction processing of event information,and they are very suitable for the high-speed AER vision system applications.