What determines the connectivity bandwidth needed in a machine vision system?

This article reviews resolution, definition, frame rate, and bit depth and how they impact connectivity bandwidth requirements. It then looks at a practical example of the requirements for an effective automotive pedestrian imaging system, including the importance of the Johnson criteria for night vision.

Pixels, their sizes, numbers, and composition are the bedrock upon which images are built. Bit depth refers to the amount of color information stored and is a key consideration that drives image quality and connectivity bandwidth needs.

Machine vision doesn’t necessarily need the high bit depth required to produce visually realistic images that people appreciate. For example, an 8-bit image can store 256 colors and is adequate for some machine vision applications, while a 24-bit image can store over 16 million colors and produces a more nuanced image (Figure 1):

24-bit color: 2²⁴ = 16,777,216 colors
16-bit color: 2¹⁶ = 65,536 colors
8-bit color: 2⁸ = 256 colors
1-bit color: 2¹ = 2 colors

bandwidth — Figure 1. 16-bit colors can deliver much more nuanced images compared with 8-bit colors (Image: *Cloudinary*)

Resolution

Like bit depth, resolution is a key specification. It’s the number of pixels that a camera can capture in a single image or a single frame for video. High-resolution cameras can capture finer details using smaller pixels. However, smaller pixels can be susceptible to noise. Getting the required image quality requires balancing the resolution and pixel size tradeoffs. It generally takes at least 2 pixels to capture an image’s most minor feature of interest.

Resolution is commonly expressed as the total number of megapixels (MP) in an image rounded off to the closest million or as the number of pixels in the horizontal dimension. For example, a Full HD image has 1920 × 1080 pixels. The resolution of these cameras is referred to as 2MP or 1080p.

Higher resolution means larger file sizes, more data to transfer, and higher bandwidth needs. It also demands more processing power.

Definition

The definition is often confused with resolution. They are related but not the same. The definition refers to image clarity. More pixels (higher resolution) can contribute to clarity. Various compression algorithms are often applied to images to reduce the file size. They don’t reduce the resolution but can result in lower definition (maybe by reducing the bit depth) and a blurry image.

Frame rate

The frame rate is the number of images captured per second, measured in frames per second (fps). A higher frame rate captures motion smoothly but requires more processing power and bandwidth and involves less exposure time. It also often requires a high-quality and more expensive image sensor.

A high frame rate (HRF) camera captures images faster than the 24 fps standard originally developed for movies shown in theatres. HFR cameras can be important in automotive safety applications, autonomous robots, industrial inspection processes, and accurately capturing barcodes and other markings on objects moving along production lines.

Resolution and frame rate tend to limit each other. The total bandwidth required can be estimated by multiplying the resolution in pixels per frame by the frame rate in fps, producing a measure of the pixels per second. That means a tradeoff between image quality and speed must be balanced to arrive at an optimal solution for a machine vision system.

Johnson criteria and neural networks

The Johnson criteria relate to object detection, recognition, and identification (DRI) in machine vision systems. They are based on the number of pixels needed to evaluate an object or a person accurately.

The images in Figure 2 measure pixels per meter (ppm) for DRI. The imaging system must produce enough ppm at the location of the most distant object (in this case, a pedestrian) to be perceived to enable a convolutional neural network (CNN) to reliably identify a pedestrian, including their position in the frame and a distance away from the imager. The minimum required resolution for Classification, between Recognition and Identification, is 10 ppm. That means a 0.3 MP video graphics array (VGA) imager with 640 pixels horizontally has a horizontal field of view (HFOV) of 64 meters.