基于点目标连通域标记的实时特征提取及其分布式运算

Real-Time Point Feature Extraction Based on Connected Components Labeling and Distributed Computing

在高速运动目标的视觉测量中,高分辨率、高帧频的图像序列带来了大量待处理数据,如何快速地从这些数据中识别合作目标并提取其特征信息,成为高速视觉测量中的难题。对此,针对高速相机每个时钟周期多像素并行传输的数据特点,提出一种基于多维金字塔的硬件加速处理结构,实现连通域的全局搜索与标记,并根据标记结果完成对应特征的实时提取。在现场可编程逻辑门阵列中,通过金字塔结构的二维处理节点阵列与多标签特征统计的一维处理阵列,将数据流的高密度运算均衡分布于各运算节点,结合流水线并行,将视觉系统中占据较高耗时的全局搜索与标记过程在图像传输中同步实现。通过功能验证与实时性分析,该特征提取的分布式运算结构可在Camera Link接口85 MHz的时钟频率下,实现680 Mpixel/s的数据流实时处理,可作为预处理部分应用于高速视觉测量系统。

In the vision measurement for high speed moving target, the image sequence with high resolution and high frame rate brings a lot of datas. How to recognize the cooperation target quickly and extract the feature information in real-time is becoming a challenge in vision measurement system. In each clock cycle of high speed camera, multi-pixel is transmitted in parallel. In this situation, a hardware accelerating structure based on multi-dimensional pyramid is proposed, the task of global searching and connected components labeling is realized, the features of connected components are extracted real-timely based on the labeling results. In field programmable gate array, with the pyramid structure formed by two-dimensional processing element （PE） array and multi-label feature statistical structure formed by one-dimensional PE array, the high density computing process of data flow is distributed into each PE node equally. Combining with the parallel pipeline, the global searching and labeling process which takes a remarkable time-consuming in ordinary vision system is accomplished synchronously with image transmission. It is verified that the distributed computing structure for feature extraction can deal with real-time data processing at the flow rate of 680 Mpixel/s at the frequency of 85 MHz with CameraLink interface. As a preprocess part, it can be used in high-speed vision measurement system.