Developments in machine vision camera interfaces

The choice of camera data interface is a vital consideration both in traditional PC-based machine vision systems and in the increasingly important area of embedded vision. Higher and higher resolution CMOS image sensors and increasing inspection speeds require large amounts of data to be transferred between the camera and processor.

The bandwidth needed for the application and the image data transmission distance are critical requirements, although other factors such as triggering accuracy and latency in the system are also important.

A number of data transfer hardware interfaces have been developed specifically for the machine vision sector over a number of years, including CameraLink, CameraLink HS, GigE Vision, USB3 Vision, CoaXPress (Table 1).

Table 1. Camera interface performance parameters

Interface Image Data Throughput Cable length (copper) Frame Grabber
GigE Vision 115 MB/s 100 m No
CameraLink Up to 850 MB/s 4m-10m* Yes
USB3 Vision 3-400 MB/s 3-5 m No
CameraLink HS 2100/3300 MB/s 15 m Yes
CXP-6 625 MB/s 40 m Yes
CXP-6 x4 2500 MB/s 40 m Yes
CXP-12 x4 7200 MB/s 25 m Yes

*Dependent on configuration and clock speed

These all rely on the GenICam standard to provide generic software interfaces to ensure complete plug & play functionality for imaging applications. Each standard will require the use of particular cable types, connectors and other components, which form an integral part of the vision system structure. Changing the infrastructure once a particular interface has been implemented can be difficult and costly.

Thus recent developments utilizing the GigE Vision platform (2.5GigE, 5GigE and 10GigE) and the CoaXPress platform (CXP 2.0) enable performance enhancements to be implemented without major changes to the system. For embedded vision applications, USB3 Vision has been the most commonly used with the cost-effective MIPI CSI-2 interfaces growing in adoption.

Building on GigE Vision

The GigE Vision standard is managed by the AIA trade association. It allows the use of existing low cost Ethernet cables, connectors, switches and other components for transmitting image data over distances up to 100 metres using copper cable and more with switches or fibre adapters.

Gigabit Ethernet also offers the potential for creating different implementation models and complex networking topologies. These include inspection ‘networks’ split up into different functional zones, but all controlled by a single workstation (Figure 1).

Figure 1

Figure 2

Another option is to use existing network structures to transmit data from remote locations to any number of different workstations. As each camera is independently locatable on a network via its IP address, it can be viewed, controlled and monitored from any PC on the network (Figure 2).

Multicast connectivity opens up scalable solutions with image data simultaneously transmitted to multiple processing PCs. GigE Vision 2.1 also includes improved real-time synchronization of multi-camera systems utilizing the IEEE 1588 Precision Time Protocol and features multi-part transmission which enables the transmission of more complex data structures used in 3D imaging or any application which would benefit from a 3 coordinate data structure.

While GigE Vision offers undoubted flexibility, the maximum bandwidth of 115 MB/s has been a limiting factor for some applications. Individual camera manufacturers have offered increased bandwidth based on parallel cabling configurations (LAG) or proprietary software approaches, however the introduction of NBASE-T technology has significantly enhanced the bandwidth within the GigE Vision framework.

NBASE-T technology from the NBASE-T Alliance™, is an extension to the IEEE 802.3 Ethernet standard and increases data transmission up to speeds of 2.5, 5 and 10 Gbit/s for 2.5BASE-T (2.5GigE), 5BASE-T (5GigE) and 10BASE-T (10GigE) respectively (Table 2).

Table 2. NBASE-T interface performance

Interface Image Data Throughput Cable length (Cat5e) Max. cable length (Cat6a/Cat 7)
GigE Vision 115 MB/s 100 m 100 m
2.5GigE 250 MB/s 100 m 100 m
5GigE 570 MB/s 100 m 100 m
10GigE 1100 MB/s 55 m 100 m

Since the Ethernet networking stack is divided into a number of different layers which are isolated from each other, camera manufacturers have been able to create 2.5GigE, 5GigE and 10GigE interface solutions that communicate using the GigE Vision standard, meaning that the technology is future proofed.

Virtually any modern PC can be upgraded for these standards using a relatively inexpensive Network Interface Card (and driver) with GigE Vision-compliant software also being compatible with NBASE-T.

Data transmission distances of 100 m are possible for 2.5GigE and 5GigE and 55 m for 10GigE using Cat 5e cable. Cat 6a and Cat 7 cable enable 100 m transmission for all 3 platforms or significantly more when using fibre adaptors and configurations.

Since using 10GigE at full speed on copper cable can generate a lot of heat as 10GigE chipsets have a high power consumption, other system level challenges need to be considered if the full speed of the interface is required.

Going for speed with CoaXPress

While the introduction of NBASE-T technology has significantly increased the bandwidth for GigE Vision systems, many applications require even more. For those that also need long data transmission distances, the CoaXPress standard, administered by the Japan Industrial Imaging Association, offers considerably greater bandwidth, while utilising coaxial cable which is used extensively in industrial, medical and defence applications.

CoaXPress is an asymmetric high speed point to point serial communication standard which requires a frame grabber and is scalable over single or multiple coaxial cables.

A single cable can transmit up to 6.25 Gbit/s from the camera to the frame grabber over distances up to 40 metres which is equivalent to approximately 5 or 6 times the standard GigE bandwidth and exceeds even 5GigE. This can be further increased to 25 Gbit/s using 4 parallel cables, or lanes (Figure 3).

Figure 3

However in 2019, CXP 2.0 was released, which effectively doubles the lane performance compared to CXP 1.1, giving up to 12 Gbit/s using a single lane (CXP-12) and 50 Gbit/s for a 4 lane system (Table 3).

Table 3. CXP 2.0 transmission distance examples

Interface Image Data Throughput Cable length (copper) Frame Grabber
CXP-12 1250 MB/s 30 m Yes
CXP-6 625 MB/s 60 m Yes
CXP-12 x 4 5000 MB/s 30 m Yes
CXP-6 x 4 2500 MB/s 60 m Yes

For applications requiring faster bandwidth than 10GigE with long data transmission distances, the new CXP 2.0 CoaXPress standard starts where 10GigE finishes. CXP 2.0 also brings higher trigger rates and improvements in cable length and data sharing as well as an increase in reliability and reporting.

CXP 1.1 uses BNC or DIN 1.0/2.3 connectors, while CXP 2.0 requires micro BNC connectors. Both CXP 1.1 and CXP 2.0 cameras are compatible with CXP 1.1 and CXP 2.0 frame grabbers, but CXP 2.0 performance is only achievable when both the frame grabber and the camera support the standard.

Data, control, real-time trigger and power are all provided over the same cable. CXP 2.0 also introduces a variety of operational technologies including data that can be simultaneously streamed to more than one host and up to 4 cameras can be connected to a single frame grabber.

The 40 metres transmission distance is extended to 60 metres for CXP 2.0 operating at the 6.25 Gbit/s bandwidth normally associated with CXP 1.1.

MIPI and CSI interfaces - embedded systems

For embedded vision systems bandwidth is just as important as for PC systems but transmission distances are generally much less of an issue as the sensor and processor tend to be closely coupled. Since embedded vision solutions generally incur high development costs, they are best suited for high volume applications where component costs can be kept low and the development costs spread over many units.

The two camera data transmission interfaces most widely used in embedded systems are MIPI CSI-2 and USB 3.1. MIPI CSI-2 is a specification of the Mobile Industry Processor Interface (MIPI) alliance (a global organization serving industries that develop mobile and mobile influenced devices) and is used extensively in mobile devices such as smart phones and tablets to connect the sensor to the processor but is not part of a machine vision standard interface.

Many embedded processors include a direct physical MIPI camera interface. The CSI-2 protocol contains transport and application layers and natively supports D-PHY and C-PHY. CSI-2 D-PHY offers up to 2.5 Gbit/s per lane with 4 lanes up to 10 Gbit/s compared to 5 Gbit/s for USB 3.0/3.1. Other factors such as the integration effort required and the load on the host (Table 4) are also important considerations for embedded applications.

There is an increasing availability of the MIPI CSI-2 interface on embedded processor boards and both board-level and housed cameras featuring MIPI CSI-2 are available.

Table 4. Interfaces for embedded vision

  CSI-2 D-PHY USB3 Vision GigE Vision
Bandwidth Highest High Medium
Cable length Up to 0.6 m Up to 8.0 m Up to 100 m
Control Difficult Easy Easy
CPU load on host Low Medium High
Availability on embedded processor boards High Medium Medium
Software Complex Easy Easy
Cost Low Medium Medium
Size Very small Small Medium

Continuing development

As the ability to improve vision systems based on existing infrastructure is key in keeping costs down, the availability of NBASE-T and CXP 2.0 interfaces take on even greater importance.

The Ethernet Technology Consortium, established to develop 25 Gbit/s and faster Ethernet specifications, have developed a low-latency forward error correction (FEC) specification for 50 Gbit/s, 100 Gbit/s and 200 Gbit/s Ethernet networks which could eventually pave the way for higher bandwidth GigE Vision systems.

For embedded systems GenICam transport layer support for MIPI CSI-2 is getting close to availability making integration of these lowest cost processor – camera modules easier to develop and will have a positive effect on the time to market for embedded vision applications. In addition, the USB4 specification was published in late 2019 but neither interfaces nor cameras are yet available.

Interfaces and image acquisition software

Since its inception in 1997, one of STEMMER IMAGING’s Common Vision Blox central goals has been to provide a simple single API that supports any acquisition device across multiple interfaces and manufacturers.

We embraced standardisation being directly involved in the formation of GenICam, GigE Vision and USB3 Vision standards and always take all standards into account in the development of acquisition engines for our independent acquisition platform.

Based on a well defined architecture and consistent modularity, Common Vision Blox enables manufacturer-independent exchange of acquisition hardware and technology at any time. The perfectly designed model for image acquisition and the driver structure enable complete decoupling of the algorithm from the image acquisition.

Related products for the camera interfaces