We use cookies to personalise site content, social media features and to analyse our traffic. We also share information about your use of this site with our advertising and social media partners.
10 minutes, 34 seconds
-445 Views 0 Comments 0 Likes 0 Reviews
Today, machine vision is used in a huge number of industries and cases, and the growth rate of this particular segment is leading among all areas of artificial intelligence. The effects and benefits that it brings allow us to make predictions that in the next five years the machine vision market can grow five times. Today we analyze the current level of demand for machine vision, assessed future prospects and explain in which cases the compatibility of machine vision with other video surveillance systems is appropriate.
Interest in machine vision is growing in many areas. First of all, the clarity of the problem statement is important for him:
There are already many providers of algorithms and solutions for mass objects (people, cars, text) on the market. For unique objects, it may be necessary to develop a detector from scratch, which is time and money consuming.
The boundary between these kinds of objects is fuzzy, and more and more detectors are emerging for new applications. Thus, the number of tasks for machine vision is constantly growing.
The possibility of obtaining high-quality images for machine processing. A wide variety of sensors, mainly with a global shutter and a pixel size of 2.5 microns. Standard interfaces for optics, a rich set of functions for managing shooting modes, standard interfaces for connecting to computers, built-in and remote, for image processing and recording.
Machine vision requires a match between the characteristics of the cameras and the intended objects for recognition in the intended operating conditions. If there is a match, then the problem can be solved. For machine vision, it is better to take IP cameras right away so that the video stream is digital, and then select the characteristics for the task. For video surveillance systems, cameras with simpler characteristics were enough, since a person was involved in interpreting the video.
With the development of matrix sensors, the range of tasks solved by matrix cameras has noticeably increased and included, among other things, some of the tasks previously solved using linear cameras. The absence of the need to use an encoder, as is the case with line cameras, also speaks in favor of matrix cameras.
Line cameras are still a good choice for visual inspection of roll materials. However, the latest generations of matrix camera sensors practically deprive line cameras of their main advantage – higher sensitivity due to the larger pixel size. Therefore, in many tasks, it becomes preferable to use matrix cameras on sensors of the latest generation due to the obviously better price, high image quality of moving objects even at low exposures, and rich functionality.
Linear sensors can be in demand for machine vision in industrial production with the tasks of analyzing the quality of extended moving objects – pipes, rolled metal, etc. Outside this segment, it is more convenient to work with cameras with conventional rectangular sensors.
Machine vision cameras are built into conventional video surveillance systems to solve additional tasks – identifying people or vehicles. Machine vision cameras require the use of networks with a high data transfer rate (GigE or USB 3.0), so they can only be embedded with video stream compression devices if the network does not provide the required data transfer rate. Otherwise, it is advisable to perform recognition on a processor directly connected to the camera and transmit only the recognition result.
Since the nature of machine vision cameras is fundamentally different (the processing of the original images is assumed, and not their transmission in some standard format, as is inherent in IP cameras), then we should talk about the compatibility of rather software and hardware solutions made on the basis of machine vision cameras. vision. Most often, such solutions are presented in the form of a combination of a camera or cameras and an embedded platform, which must be integrated with video surveillance systems. At the same time, the embedded platform should be perceived by the system as a kind of “supercamera” – not duplicating (this makes no sense), but providing additional functionality due to its capabilities, which is inaccessible to ordinary, even intelligent, IP cameras.
If it is planned to develop a security video surveillance system to the tasks of automated recognition, then compatibility is a must.
Image pre-processing is already possible in machine vision cameras, which is a combination of sharpening, noise reduction, color smoothing and 5×5 debayerization. These functions are useful in almost any application and do not create any additional load on the CPU. If it is necessary to use more complex computational algorithms, it is reasonable to use a powerful compact single-board computer, which is now one of the main trends in machine vision.
Continuing the idea of the inexpediency of machine vision cameras duplicating the functionality of IP cameras, including those with built-in video analytics, the processing of images received from a camera, sometimes from several cameras (for example, at an intersection), should be performed on a separate computer, which can, and in some cases and should be placed next to the camera. First of all, we are talking about tasks with high requirements for image quality, which is inaccessible to standard cameras due to obvious market limitations of their cost and, as a result, the technical level, the need to use additional information from other sensors, systems and / or cameras, lidars and other equipment. .
It is important to distinguish between video signal processing for general image enhancement (sharpness, illumination, blur reduction) and machine vision processing (object and event detection). These are different tasks that are performed by different software and usually on different processors. There are situations when it is beneficial to place machine vision processing near the camera (for example, recognition at the entrance to the office, recognition of numbers and speed of cars, a counter of the number of customers in the queue). In these cases, the task is relatively simple, and delays or failures in the network connection can make the solution unacceptable, or you want to avoid overloading the network channel with video streams from a large number of cameras.
Integrating machine vision algorithms and calculators into a camera body looks attractive, but is still a less convenient option for most tasks.
The calculator located outside the chamber leaves more flexibility for updating the calculator itself and its software without changing the chamber. This is especially valuable for machine vision tasks that are in development (for example, you can already recognize people and count them, then you need to recognize whether these people are wearing helmets and vests at a construction site, then you need to see medical masks on them, etc.).
