After my articles and videos about 3D cameras — one of the most popular questions was — “Which 2D camera should I choose for my project?”
This question is orders of magnitude more complicated than it seems. For most systems that I have seen, “choosing a camera” is part of the final product — as important and complicated as “collecting a dataset” or “training a neural network.”
Link to HBVCAM
Why?
Today, it’s pretty clear that expertise in VLM and Edge models differs. The same is true of cameras. When people develop low-power cameras for home use and cameras for astronomy, they will use different approaches and solutions.
So, I’ll warn you right away — this guide does not pretend to be comprehensive. I tried to focus on a few things:
Also, when preparing the article and video, I asked a few experts to criticize and add their point of view on these issues:
By the way. This article also has a video on YouTube. In this video I tell a few additional examples and share practice experience:
When choosing a camera, you need to evaluate several characteristics at a high level:
To estimate this, let’s check all these characteristics on the camera
USB, CSI, and LAN are the primary interfaces that account for 95% of cameras. We will talk about them in detail later, and here are a few words about the rest. Partly, these are wrappers over existing ones:
Each interface has two main parameters: speed and signal delay. The slowest ones now are old cameras (USB 2.0, etc.). The fastest ones utilize several PCIe lines.
The further your camera is from your processing unit, the bigger the difference. The more intermediate processing your signal will go through — the later you will output your result.
We can talk about optics endlessly. But let’s focus on the main characteristics:
Main characteristics of the matrix:
Also, some cameras have their own light system; sometimes, you need to choose whether to build your system or use existing boxing.
Also, for some applications, it’s super important to choose a shutter type. Global vs. Rolling Shutter. Global shutter allows you to have the same moment of exposure for all pixels. Also, for some cameras mechanical shutter, it’s still an option(astronomy, for example). But it’s quite rare.
Some cameras have an internal CPU/NPU/GPU for image processing, and for some, you need to use another device.
The first advantage is that preprocessing works on the camera. Some USB cameras can stream a raw image, but most cameras will stream already processed and compressed images. This allows you to utilize the channel width optimally.
The second advantage is that most USB cameras are “plug and play.” They will work out of the box under Windows andmost Linux systems, MacOS, Android, etc.
The third advantage is availability. A USB camera will be sold in a nearby store. Of course, not every USB camera is so easy to buy, but this will be reasonable for most examples.
The fourth advantage is that USB 3 is a fairly fast protocol. Especially considering compression.
The first disadvantage is the limited cable length. The camera out of the box will most likely have:
The second minus. Simple cameras can have a large delay and it can be unstable.
The third minus. Sometimes, finding a ready-made camera with the target characteristics isn’t easy.
The fourth minus is that drivers can be unstable. Recently, this part has improved. But 6–7 years ago, the driver could freeze on Jetson, requiring a system reboot.
In my opinion, USB cameras are super good for prototyping. They allow you to quickly connect something and minimize problems caused by drivers and codecs.
Very often, small-scale indoor solutions/factory solutions are made on USB cameras, where the camera’s price is no longer a determining factor at $200-$300.
USB cameras are good for providing high FPS with high-quality images.
In general, USB is when it is “Fast” and “Good” but “Expensive”.
Several examples of USB cameras that can be used in projects.
First option. Cheap, quick to buy, good quality for student projects, testing algorithms, etc. — Logitech C270 (Logitech C920/C922 has a little bigger resolution and framerate)
Second option. Expensive, rich, good quality, lots of implemented interfaces, good drivers and lenses — Basler cameras. We used them a lot back in –. They where super nice.
Third option. Good Chinese cameras — Vision Datum. In general, covers most of what Basler can do. But the software is a little worse, the design is a little worse. Etc., etc.
Forth Option. Super Small USB Boards (1, 2). They are nice if you want build your own camera.
The most configurable protocol is CSI. You can use a different number of lines or a different approach to transmission. Both the fastest and the slowest cameras are made with CSI protocol.
Cameras transmit raw information — this allows you to preprocess the image yourself. Also, it’s one of the fastest protocols.
There are a lot of camera vendors. You can choose a camera that suits the task.
Also, there are a lot of different connectors based on the top of the CSI protocol.
Limited length. This can be partially fixed with coaxial connectors, but this increases the system’s complexity and cost. People try to stay within the 30cm main range for most everyday applications. But it’s hard to find something much longer then 10m.
If your board does not have a separate chip for image encoding, you must do it on the processor. Which will additionally load it
More development time if you go down to the lower level.
CSI cameras are different. They require mutual calibration of the camera and the board. If this is some “arbitrary camera and arbitrary board,” — they may not work out of the box.
When you need to maximize control over the system, when you need to achieve the minimum production cost but do not want to lay out your own matrix yet, or when the final device should be small (this can be achieved with USB, but with CSI, the choice will be greater).
Also, if you are still developing your own board — using CSI is a fairly logical step.
Of course, there are a few classic examples that everyone uses
RPi camera — (works on a lot of boards)
On the Nvidia website, you can find a lot of cameras that work with Jetson.
And here is, for example, another extensive list of different cameras.
A clear protocol that will work with all devices. Problems arise much less often than with USB. Lots of goodies out of the box: the camera can be on the other side of the world. Multicast.
There are many cameras on the market. You can find good lenses, good lighting, and competent filters right out of the box.
You can find good cases. Dust and moisture are protected with heating and anti-fogging.
You can find full control: rotation, zoom, pointing, and sharpness.
High latency — much higher than USB/CSI. Usually much more expensive than USB and CSI. Cheap models have almost no settings and may not have a good picture.
You are essentially buying a box. No ability to configure internal algorithms. Unless you are developing a camera.
You need a network infrastructure. Most likely, you will need a router and lay networks.
When latency is not important to you. When there are many users. When you have already installed the camera and the algorithm needs to be run on a ready camera. When you have a central computer where you want to calculate everything. When the working conditions are challenging.
From expensive cameras, Moxa and Axis are worth mentioning.
Middle class — HikVision and Dahua.
You can go to Aliexpress for cheap cameras. But remember that for some, even receiving a stream is a problem. The cameras I chose 2–3 years ago are already out of stock. Different cameras from the same vendor may not work, etc.
There is no straightforward solution to the question, “What camera should I choose?” Even top companies experiment a lot.
So, the best approach is to schedule an additional amount of time to research this question for your case as well. After this:
While you wait for the camera of your dreams, take some simple USB/LAN cameras for experiments.
For more USB Global Shutter Camera Moduleinformation, please contact us. We will provide professional answers.
CMOS (Complementary Metal-Oxide-Semiconductor) sensor captures light and converts it into electrical signals. It employs standard semiconductor manufacturing techniques, providing a cost-effective and energy-efficient alternative to CCD sensors. They achieve higher frame rates and lower power consumption.
To explore why CMOS has already surpassed CCD, please visit CMOS sensors vs. CCD sensors: why are CMOS sensors ruling the world of embedded vision? - e-con Systems
The Exmor R series uses BSI technology, enhancing sensitivity. Its sensitivity is approximately twice that of a front-illuminated sensor.
To learn more about how Sony Exmor R compares with other Sony sensors, please visit Sony Exmor vs Sony STARVIS sensors; a detailed comparison - e-con Systems
Sony STARVIS sensors feature Backside-illuminated (BSI) technology for surveillance camera applications. It contributes to the creation of high-quality images in the visible and near-infrared (NIR) light spectrums with a sensitivity of mV.
However, Sony Pregius is also a series of CMOS image sensors. It has a global shutter pixel technology based on Sony's low-noise CCD construction.
To understand the revolution in sensor technology, please visit Sony STARVIS vs. Sony Pregius: The ultimate image sensor comparison - e-con Systems
When selecting a CMOS sensor size for embedded vision applications, consider the following factors :
Resolution : Sensors with larger pixel sizes achieve higher resolution.
High Frame Rate and Global Shutter : Important for applications like license plate recognition, robotic vision, drones, and AMRs.
Sensitivity : Larger sensors with larger pixels increase sensitivity.
Image Circle Diameter : Must fit or exceed the sensor size.
Low-light Performance : Larger sensors capture better low-light images.
Trigger mode allows the camera's image sensor to capture an image in response to a specific triggering action, such as the passing of an object, a strobe light flash, or the press of a button.
To know more about triggering modes of USB 3.0 camera, please visit Trigger Modes available in See3CAMs (USB 3.0 Cameras) - e-con Systems
The global shutter image sensor exposes all the pixels in the sensor to light simultaneously within the programmed exposure duration during each image acquisition cycle and reads out simultaneously.
To know about the right imaging solution for moving targets, watch Global Shutter vs Rolling Shutter - eVision Hub on YouTube
Capturing fast-moving objects : The global shutter function captures high-speed motion without motion artifacts in time, as the exposure happens at the same time across the sensor.
Improved Synchronization : It ensures uniform synchronization since there is no time difference across the image, which is crucial for industrial inspection, 3D scanning, etc.
High Reliability : In environments with vibration or rapid movement, it offers constant image quality without artifacts and motion blur from sliding shutters.
Smart parking : Improved safety with accurate real-time imaging for driver assistance systems.
Industrial automation : Realizing precise and accurate inspection of fast-moving production lines.
Warehouse robot : Enhanced NIR lighting sensitivity, which makes it ideal for package tracking under red lighting.
Drone : Achieving clear and undistorted images while moving, enabling them to perform tasks such as 3D mapping, object tracking, and real-time image analysis more effectively.
Live cell imaging : Capturing accurate, high-resolution images of Rapid cellular movement without distortion.
Watch our joint webinar with industry leaders NVIDIA and Connect Tech on Next-Gen Autonomous Mobility Systems on YouTube, to discover how global shutter sensors are utilized in various autonomous systems.
The rolling shutter image sensor reads the pixel row by row when exposed, with a temporal offset between each row. The first row of pixels begins capturing light, followed by the remaining rows at a fixed time offset, which leads to a temporal delay between the start and end times of light collection for each subsequent row. Such temporal offset influences image quality, often resulting in artifacts like skew or wobble.
No Parasitic Signal : Rolling shutter sensors don't have any parasitic signal that generally gets accumulated due to pixel architecture in a global shutter sensor.
Lower sensor cost : The simpler architecture makes the sensor easier and less expensive to manufacture, resulting in lower production costs.
Higher light sensitivity : Achieves the same sensitivity in a much smaller form factor than a global shutter sensor.
The rolling shutter effect occurs when an object moves faster than the image sensor's exposure and readout time, causing distortion due to sequential scanning of sensor pixels.
To understand how to overcome rolling shutter artifacts, please visit Differences between rolling shutter artifacts and motion blur - e-con Systems
Digital Microscopes : For indoor usage, a rolling shutter image sensor is ideal because of its high sensitivity and high resolution, which enables it to take high-quality images in low-light conditions
Automatic Number Plate Recognition : The rolling shutter image sensor is featured with GRR mode – capturing images of vehicles on the road—from fast-moving vehicles that break violations to slow-moving ones that enter and exit parking lots.
Next-generation Ultra HD Webcams : The rolling shutter image sensor provides clear images in low light conditions by preserving details and minimizing noise in the output image even at 0.1 lux.
Dentistry : Achieving high-resolution and high-quality videos and images from dental loupes.
Global Reset Release (GRR) mode is a unique feature offered by some rolling shutter sensors to overcome rolling shutter artifacts. In the GRR mode, all the rows are exposed to the scene simultaneously. This feature enables the rolling shutter sensors to simulate the characteristic of a global shutter sensor during the image acquisition of moving objects.
To know how to simulate global shutter behavior with GRR Mode of rolling shutter sensor, please visit What is Global Reset Release? How does it work in e-con cameras? - e-con Systems
Monochrome sensors have higher quantum efficiency and better sensitivity, resulting in higher output quality, especially in low-light conditions. Color cameras have color filter arrays that limit the number of photons reaching the sensor, which is absent in monochrome sensors.
To get insights on why monochrome cameras are more advantageous than color cameras in low lighting conditions, please visit A short guide to why monochrome cameras have the edge over color cameras - e-con Systems.
For low-light applications requiring clear images at night, consider a camera with sensors specifically designed for high sensitivity, such as Sony's STARVIS technology, which is known for excellent performance in low-light scenarios, providing clear images with minimal noise.
Want more information on 2mega Pixel USB Cameras? Feel free to contact us.