Composition, key factors and challenges of visual positioning system

Machine vision is mainly used in manufacturing for visual guidance, dimension measurement, product inspection, object recognition and other fields. In these fields, one of the most basic algorithms is product identification and positioning. For example, a visually guided robot must identify the product to be grasped in the image and locate the coordinates before guiding the robot to the product location. The same is true for dimension measurement and product inspection. Before measurement and inspection, we must first know whether there is a product and where the product is located before applying various subsequent analysis tools. Therefore, product identification and positioning is a basic problem.

01 Composition of the visual positioning system

The robot positioning system based on machine vision includes a camera system and a control system. The camera system includes a computer (with an image acquisition card) and a camera, which mainly collects visual images and applies machine vision algorithms. The control system includes a control box and a computer to control the specific position of the computer terminal. The work area is photographed using a CCD camera, and the computer is used to identify the image, obtain tracking features, complete data calculation and recognition, and use the inverse kinematics method to obtain the error of each position of the robot, and then control the high-precision terminal execution module to scientifically adjust the position and posture of the robot.

02Key factors of visual positioning system

In the field of industrial production, especially in the application of industrial robots, visual recognition and positioning systems are particularly important. In actual production, we should not only pay attention to whether we can accurately grasp, but also pay attention to its speed. This has always been a problem in the industry. The industrial robots we often encounter are usually slow to grasp. Once the speed is increased, the accuracy of grasping will be a problem. This is also a difficult problem for visual recognition and positioning systems. Let's follow Xiaoju to learn about it. The first is the amount of data. In a more complex production environment, the system needs to accurately find the products that need to be identified and positioned; the second is speed. How to increase the speed to the ms level in some standard production lines. Although the previous algorithms can play a role in ordinary production, with the continuous development of algorithms, deep learning algorithms often need to be equipped with more ideal GPUs to achieve; then there is the core of the problem, positioning accuracy. In the deep learning system, the images we see are all scaled to a certain extent. We need the entire system to match the original image with pixel accuracy; the rest is the accuracy of recognition. In many cases, we can get very little learning data. In this state, how to further improve the accuracy of recognition!

03Challenges in visual positioning

If you want to design a feasible product identification and positioning algorithm, you need to overcome several difficulties:

1. Quickly specify products Industrial products vary greatly. Therefore, for each specific application, you need to quickly specify the product you need to find from several or even one image. For example, if you need to locate the position of rivets on the current production line, you can take a photo and learn accordingly, and then search and locate it in subsequent images.

2. Quickly search for products. For a 2-megapixel image, it is usually required to identify and locate the product within tens of milliseconds.

3. High-precision positioning Industrial production has strict requirements on accuracy and tolerance, so the positioning of products must be accurate. It is now generally required that the recognition and positioning algorithm can achieve a pixel-level positioning accuracy, or even a sub-pixel level.

4. Adapt to the impact of missing, blocked, dirty, etc. If a product is blocked, resulting in a certain percentage of the product missing in the image, it is still necessary to be able to identify and locate the object. Conversely, if the surface of the product is dirty, resulting in changes in the surface features, it is still necessary to be able to identify and locate.

5. It can adapt to the influence of uneven light brightness. If the brightness of the product changes, for example, half is bright and half is dark, it can still be identified and positioned.

6. Can identify rotating products. Products can usually rotate within 360 degrees.

7. Multiple products can be identified. There may be multiple products in one image, which need to be identified and located separately.

8. Can accurately identify nearly symmetrical objects. Nearly symmetrical objects are easily identified as the wrong direction, so corresponding design is required.

9. It can cope with the polarity reversal of objects. For example, the product being studied has black text on a white background, but in fact the product image may have white text on a black background, which needs to be recognizable.