Read computer vision and robot vision

There are many similarities between computer vision and robot vision. The basic theoretical framework, underlying theory, and algorithms are similar, but the ultimate goal of computer vision and robot vision research is different: the former mainly studies visual inspection, and has high precision requirements. Speed ​​is not a major consideration; robot vision mainly studies the role of robots in the environment under visual guidance, and has real-time requirements.

1. The concept of computer vision

Computer vision is to replace the visual organs with various imaging systems as input-sensitive means, and the computer replaces the brain to complete the processing and interpretation. The ultimate goal of computer vision is to enable computers to visually observe and understand the world like humans, with the ability to adapt to the environment. But before the ultimate goal is achieved, the medium-term goal of people's efforts is to create a visual system that can perform certain tasks based on a certain degree of intelligence of visual sensitivity and feedback. (The point to be pointed out here is that computers in computer vision systems act instead of human brains, but it does not mean that computers must perform visual information processing in a human visual way. Computer vision can and should be based on the characteristics of computer systems. Processing and guidance of visual information.)

2. The development of computer vision

Visual research was based on two-dimensional before Roberts, and most of them used pattern recognition to complete the classification work. Roberts first successfully used the program to explain the world of 3D building blocks, and in a similar study later, Huffman. Clowes and Waltz et al. studied the building block world and solved the problems of explaining scenes and dealing with shadows by line segments. The study of the building block world reflects some of the characteristics of early visual research, starting with a simplified world. These efforts have contributed to the development of visual research, but it is difficult to work with slightly more complex scenes.

In the mid-1970s, some researchers, represented by Marr, Barrow, and Tenebaum, proposed a set of visual computing theories to describe the visual process, the core of which is to restore the three-dimensional shape of the object from the image. In the theory of visual research, Marr's theory has the most profound influence. Its theory emphasizes the importance of representation and proposes to study the problem of information processing from different levels. He emphasized the importance of computational theory for computational theory and algorithm implementation. Although this framework still has incomplete aspects in the details and even in the dominant thinking, there are still many controversies in many aspects, but it is still the basic framework of computer vision research.

In the mid-to-late 1980s, with the research of mobile robots, visual research is closely combined with the introduction of spatial geometry methods and physical knowledge. The main goal is to realize the recognition and treatment of roads and obstacles. In this period, active vision research methods were introduced, distance sensors were used, and techniques such as multi-sensor fusion were adopted.

3. Problems in computer vision research

Researchers from all over the world have carried out a lot of research on the various research levels of computer vision systems according to the basic theoretical framework proposed by Marr, and proposed corresponding solutions. However, in general, these methods have some problems. Or lack of versatility, or poor anti-interference ability, or multiple solutions, the reasons are as follows: First, computer vision is an inverse problem, that is, the input image is the gray scale of the two-dimensional image, which is the geometric feature of the three-dimensional object, illumination, A function of many factors such as the surface properties of the object material, the color of the object, and camera parameters. The inverse of the above parameters by grayscale is an inverse problem, and most of these problems are non-linear, the solution of the problem is not unique, and it is extremely sensitive to errors caused by noise or discretization; another reason is the visual system of Marr The framework is a top-down, modular, one-way, data-driven architecture. In-depth study of neurophysiology shows that this structure is still far from the human visual system. The cognitive process of the biological vision system is a purposeful and active process of interacting with the outside world, not just a passive one. Reaction.

4. The concept of robot vision

The robot vision system refers to the realization of the human visual function by using a computer, that is, using a computer to realize the recognition of the objective three-dimensional world. 2. Robot vision The main research is to use computer to simulate human visual function to extract information from the image of objective things, process and understand, and finally use for actual detection, measurement and control.

The sensory part of the human visual system is the retina, which is a three-dimensional sampling system. The visible portion of the three-dimensional object is projected onto the retina, and the person performs a three-dimensional understanding of the object in accordance with two-dimensional imaging projected onto the retina. If the three-dimensional objective world to the two-dimensional projection image is regarded as a kind of positive transformation, then the machine vision system has to do the inverse transformation from the two-dimensional projection image (gray array) to the three-dimensional objective world, that is, according to This two-dimensional projection image is used to reconstruct the three-dimensional objective world.

5. Development of robot vision

The robot vision system can be divided into three generations according to its development. The function of the first generation of robot vision is generally to process the image and output the result according to the prescribed process. Such systems are generally constructed from common digital circuits and are primarily used for defect detection of flat materials. The second generation of robot vision systems typically consists of a computer, an image input device, and a result output hardware. The visual information flows in a serial manner within the machine and has a certain learning ability to adapt to various new situations. The third generation of robot vision systems is currently being developed and used internationally. Using high-speed image processing chip, parallel algorithm, with high intelligence and common adaptability, can simulate human high visual function.

6. Comparison of computer vision and robot vision

There are many similarities between computer vision and robot vision. The basic theoretical framework, underlying theory, and algorithms are similar, but the ultimate goal of computer vision and robot vision research is different: the former mainly studies visual inspection, and has high precision requirements. Speed ​​is not a major consideration; robot vision mainly studies the role of robots in the environment under visual guidance, and has real-time requirements. Therefore, robot vision research has more difficulties.

7. The method of robot vision system (this article is omitted)

8. Robot vision at home and abroad

Foreign machine vision systems are used in many fields, such as underwater robots for offshore oil exploration, seabed exploration; medical robots for medical surgery and research; space robots that help humans understand the universe; nuclear industrial robots that perform special tasks, etc. . Although there is still a certain gap between the development of machine vision in China and the world advanced level, the development of machine vision system has also achieved certain results. China's visual robot application mainly has the following purposes: to replace human beings in hazardous, harmful and harsh environments, work in a clean environment; to liberate people from dirty and heavy labor; to improve labor productivity, improve product quality, and quickly Respond to market requirements and strengthen competitiveness in the international market.

9. The main problems of robot vision

The current robot vision has the following problems:

1. How to identify the target accurately and at high speed (real time).

2. How to effectively construct and organize a reliable identification algorithm and implement it smoothly. This is expected to be a new breakthrough in high-speed array processing units, as well as algorithms (such as neural network methods, wavelet transforms, etc.), so that functions can be implemented in parallel with a high degree of computation.

3. Real-time performance is an important issue that is difficult to solve. The low image acquisition speed and the long time required for image processing bring significant time lag to the system. In addition, the introduction of visual information also significantly increases the computational complexity of the system, such as calculating the image Jacobian matrix, estimating the depth information, and so on. Image processing speed is one of the main bottlenecks affecting the real-time performance of the vision system.

4. Stability is the first consideration of all control systems. For visual control systems, both position-based, image-based or hybrid-based visual servo methods face the following problems: how to ensure system stability when the initial point is far away from the target point That is, increase the stable area and ensure global convergence; in order to avoid the servo failure, how to ensure that the feature point is always in the field of view.

10. Robot vision should be further studied

According to the current situation, robot vision should be further strengthened in the following aspects:

1. The selection of image features. The performance of visual servoing is closely dependent on the image features used. The selection of features requires not only the identification of the indicators but also the control indicators. From a control point of view, the use of redundant features can suppress the effects of noise and improve the performance of visual servoing, but it will add difficulty to image processing. So how to choose the features with the best performance, how to deal with the features and how to evaluate the features are all issues that need further study. For tasks that may sometimes need to switch from one set of features to another, consider combining global features with local features.

2. Combine the research results of computer vision and image processing to build a special software library for robot vision system.

3. Strengthen the dynamic performance of the system. Most of the current research focuses on determining the desired robot motion based on image information, but lacks research on the dynamic performance of the entire visual servo system.

4. Use the results of smart technology.

5. Take advantage of the results of active vision. Active vision is a hot topic in the field of computer vision and robot vision research today. It emphasizes the ability of the visual system to interact with its environment. Unlike traditional universal vision, active vision emphasizes two points. One is that the visual system should have the ability to actively perceive, and the other is that the visual system should be based on a certain task (TaskDirected) or purpose (PurposiveDirected). Active vision believes that in the process of visual information acquisition, the parameters of the camera, such as direction, focal length, aperture, etc., should be more actively adjusted and the camera can be quickly aimed at the object of interest. More generally, it emphasizes the gaze mechanism (AttenTIon), emphasizing the selective perception of signals distributed over different spatial extents and time periods with different resolutions, which can be used to pass camera physical parameters on the hardware layer. The adjustment implementation can also be implemented by selectively processing the obtained data on the algorithm and presentation layer based on the passive camera. At the same time, active vision believes that visual processes that are not based on any purpose are meaningless. The visual system must be associated with the purpose (such as navigation, recognition, operation, etc.) to form a perceptive/acting ring (PercepTIon/AcTIonCycle).

6. Multi-sensor fusion problem. Vision sensors have a certain range of use. If they can effectively combine other sensors and take advantage of their complementary performance, they can eliminate uncertainty and achieve more reliable and accurate results.