Design Of Image Data Processing System For AER Image Sensor

Traditional CMOS image sensor based on the sampling mode of frame,resulting in redundancy of image data.It limits the efficiency of moving object recognition.Through the research of the principle of biological vision,people put forward the AER(Address Event Representation)sensor.It has the advantages of low redundancy,fast speed and low latency,and has become the current research focus.As the most typical AER image sensor,DVS(Dynamic Vision Sensor)is especially suitable for the field of moving object recognition.In order to achieve recognition of moving object based on DVS,firstly,the image data generated by DVS is transmitted to PC in real time using the USB 3.0 high speed data transmission based on Slave FIFO mode.Secondly,this paper proposes two target shape feature extraction algorithms to extract the features of the image data.The first one is a complete target contour extraction algorithm based on DVS.In view of the deletion of contour event,two kinds of Y-event generation algorithms are designed to repair the contours of the target in horizontal and non-horizontal direction respectively,so as to extract the contour features of the targets;the second one is an adaptive target binarization method based on a dual-camera system,in which two DVSs are used to collect image data.Then the image data are processed by denoising,event mosaicking and event-to-event matching to determine the region of the target.Finally,the region feature is extracted by the image open and close operations of morphology methods,which are adopted to remove the artifacts caused by event-to-event mismatching.The image data processing system for AER image sensor is set up based on the self-developed DVS,FPGA development board,USB 3.0 development board and PC.Under 64-bit operating system of Window 7,the system can transmit the event data to the PC accurately,and the transmission rate reaches 147600 KB/s.A large number of experiments on two shape feature extraction algorithms have been carried out with MATLAB.Experimental results show that the contour extraction algorithm based on DVS can repair the deletion of contour event by 100%,and obtain a complete,one-pixel-wide contour.The adaptive target binarization method based on a dual-camera system can not only effectively extract the region features of the target,but also reduce the amount of image data to 10%~20% of the traditional algorithm.In addition,the misclassification of the proposed algorithm is only about 1% under the degradation factors of non-uniform illumination,low contrast,noise and light spots,which shows better robustness.