AER视觉传感器系统后端事件特征提取方法设计

发布时间：2018-04-26 09:16

本文选题：AER视觉系统 + 特征提取　；参考：《天津大学》2016年硕士论文

【摘要】：传统的视觉传感器以“帧扫描”为图像采集方式。随着视觉系统实际应用对于速度等性能要求的提升,传统视觉传感器遇到了数据率过大、帧频受限、动态范围低的发展瓶颈。因此,基于仿生视觉感知模型的地址-事件表达(Address Event Representation,AER)视觉传感器以其速度高、延迟小、冗余低的优势成为当前机器视觉系统领域的研究热点,该类传感器仅对发生变化的像素触发响应、异步输出稀疏表示的事件信息,从根本上消除了冗余信息的产生,特别适合定向高速物体拍摄及目标识别等机器视觉系统。本文研究了三种基于AER视觉传感器的特征提取算法,这些算法可以实时的从低冗余事件信息中提取出目标的形状特征和纹理特征,为进一步的机器识别系统研究提供数据准备。本文首先简要介绍了面阵AER视觉传感器和线阵Timed-AER视觉传感器的概念、工作原理以及基本结构,并指出该类视觉传感器所存在的事件信息不易理解、无法继承传统图像处理方法等系统缺陷。之后本文针对目标形状特征提取的需要,设计了基于AER事件对匹配的高速目标二值化方法,通过对ON/OFF事件信息进行去噪、细化、轮廓闭合等预处理获得目标轮廓的主体外形,再通过事件对匹配方法确定目标区域,完成二值化操作,实现目标与背景的分离。并设计基于等价标号思想的高速二值连通域标记方法,只对有限的事件点进行标记,避免了对全帧图像的冗余遍历,提高了标记算法的效率,实现同一视场中不同目标的标记分割。最后本文设计提出了AER卷积处理算法,通过16种Gabor模板对事件信息进行卷积,实现了事件信息不同方向、不同尺度下纹理特征的提取。通过对本文设计算法的实验分析和与传统算法的对比,仿真结果表明,本文设计的基于AER事件的目标二值化算法能够应对非均匀光照、低对比度等非理想环境条件,同时具有较高的算法效率,对于一幅512×512的图像,平均运行时间为2~4s;基于事件的二值连通域标记算法速度可以达到传统等价标号算法的1.5~8倍;而本文设计的AER卷积处理算法也能有效的提取原始事件信息在不同方向和不同尺度下的纹理特征。综上所述,本文提出的三种算法能够高效的实现事件信息的特征提取,适用于高速AER视觉系统应用领域。
[Abstract]:Traditional visual sensors take "frame scan" as image acquisition. With the improvement of speed and other performance requirements, traditional visual sensors have encountered the development bottleneck of too large data rate, limited frame frequency and low dynamic range. Therefore, the Address Event Repr based on the bionic visual perception model Esentation, AER) vision sensor has become the research hotspot in the field of machine vision system for its high speed, small delay and low redundancy. This kind of sensor only triggers response of the changing pixels, and the asynchronous output is sparse representation of event information, which fundamentally eliminates the generation of redundant information, especially suitable for directional high-speed object pat. Three kinds of feature extraction algorithms based on AER vision sensors are studied in this paper. These algorithms can extract the shape features and texture features of the target from the low redundancy event information in real time, and provide data preparation for the further research of the machine recognition system. Firstly, this paper briefly introduces the surface array A. The concept, working principle and basic structure of ER visual sensor and linear Timed-AER vision sensor, and point out that the event information of this kind of visual sensor is not easy to understand and can not inherit the defects of the traditional image processing method. After that, this paper designs the matching based on the AER event to the need of the target shape feature extraction. The high speed target two value method is used to obtain the main contour of the target contour by denoising, refining, contour closing and other preprocessing of ON/OFF event information. Then the target area is determined by the event matching method, and the two value operation is completed to separate the target from the background. The high-speed two value connected domain marking based on the equivalent label idea is set up. Method, only a limited event point is marked, which avoids the redundant traversal of the whole frame image, improves the efficiency of the labeling algorithm and realizes the segmentation of different targets in the same field of view. Finally, the AER convolution processing algorithm is designed and the event information is convoluted with 16 kinds of Gabor templates to realize the different direction of event information. The extraction of texture features under different scales. Through the experimental analysis of the design algorithm in this paper and the comparison with the traditional algorithm, the simulation results show that the target two value algorithm based on AER events designed in this paper can deal with non ideal environment conditions such as non-uniform illumination, low contrast, and also has high algorithm efficiency, for a 512 x 512. The average running time is 2~4s, and the two value connected domain labeling algorithm based on events can reach 1.5~8 times of the traditional equivalent labeling algorithm, and the AER convolution processing algorithm designed in this paper can also effectively extract the textured characteristics of the original event information in different directions and different scales. In summary, the three algorithms proposed in this paper are proposed. It can efficiently realize the feature extraction of event information and is suitable for the application field of high-speed AER vision system.

【学位授予单位】：天津大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41;TP212

【参考文献】