Research On The Technologies Of High Speed Vision Measurement System

High-speed targets can be found in many fields, such as aerospace, traffic safety, military reconnaissance, shooting test, target interception, industrial pipeline detection, biological and mechanical research. Usually, these targets moves fast, exists in diverse environment, with complex trajectory and difficult in tracking. When these targets are measured based on vision technology, the vision measurement system should with high-resolution for the demand of accuracy, and with high update-rate for better real-time ability. The big data processing task accompanied with high resolution and high frame rate become the main contradiction in high-speed vision system design. How to process the big data by the hardware with high-density computing ability in real-time become one of the problems to be solved in the measurement of high speed moving targets. The main purpose of the paper, "RESEARCH ON THE TECHNOLOGIES OF HIGH SPEED VISION MEASUREMENT SYSTEM " is to introduce large-scale parallel computing technology to the process of big data in vision measurement. With the benefit of parallel structure and fast algorithm, the problem caused by feature extraction, recognition and matching for the big data can be solved without personal computer.Through the analysis of multi-algorithm processing, both the design considerations for vision measurement system and the MIMD structure of embedded system based on modular design are proposed. With the algorithm structure optimization for the high-speed vision measurement system based on pipeline overlapping, the software structure based on system-level pipeline is accomplished. The processing chips, such as FPGA and DSP, are selected with the considerations on hardware acceleration ability, according to the demand for computing at the same time. These processing chips should provide a better performance for large-scale parallel algorithm, fixed-point/floating-point intensive computing. Data transmission has always been one of the main factors that decrease vision system real-time performance. Increase the speed in the transmission individually cannot fundamentally solve the problem of time-consuming. Focus on this problem, a data sharing strategy based on transmission pipeline is proposed, which shares the memory between different processors and decrease the time-consuming of data transmission into nanoseconds.In the image processing of vision system, the iterative step of global search or neighborhood search usually take the longest time-consuming and become the critical path of the system. In cooperation target searching, the process can be conclude into the connected region searching and feature analysis for the binary image. For this process, a distributed computing PE array based on multidimensional pyramid structure is proposed in this paper. As the system pipeline overlapping with image transmission step, this structure could labeling the connected region and extracting the feature from parallel correlated data real-timely in each transmission clock. For the labeling step, a fast labeling method based on one pass accompany with triple scan is proposed. With one-dimensional label PE array in the following step, the problem of connected region labeling, feature statistics, and features fusion for conflict label could be solved real-timely. Combining with the methods above, each PE structural model and organization mode in parallel computing is also proposed. Experiments show that the method can finish global searching and feature extracting in the limited cycle of pipeline, which presents well real-time performance and robustness.In multi-rigid-body matching step, the false target and abnormal state of the markers are recognised by the analysis for the feature vector from the marker set. So, the effective marker set in the current frame is obtained and each marker in the set is matched in the following step. In this process, each marker not only should be matched between frames, but also should be classified according to the spatial and temporal correlation. The matching process of the markers is a table-lookup process which is associated with markers’ number and harmful to the stability of the system. So two-step matching principle is proposed. The motion vector of the markers between frames is estimated by Kalman filtering, and get the first step matching based on markers’ neighborhood. Then the second step matching is performed based on the invariability of cross ratio. The analysis for the complexity of algorithm shows that two-step matching could decrease the matching times effectively when the number of markers increases in the image. In the classification step of multi-rigid-body, the angle cosine of spatial feature and motion feature is used to measure the similarity of the markers. Then, marker clustering is achieved.In the section of system verification, the correctness, robustness of feature extraction, and timing ability of hardware acceleration are verified. The influence between algorithm structure and resource usage are also analyzed. Finally,the dynamic measurement of the system and real-time performance are verified. The results show that hardware acceleration could solve real-time computing problem in the global searching and the feature extraction step of high-resolution image. Combining with the fast matching algorithm, in the condition of 5 markers, 2048*2048 resolution, 583μs exposure time, the update-rate of binocular vision measurement system could reach 137.8FPS, with the data flow speed at 1.1Gpixel/s in real-time processing. The max update-rate of the algorithm in real-time processing could reach 866.9FPS, which shows a better real-time ability.