Design Article
Embedded vision boosts driver-assistance designs
By Tom Wilson, CogniVue Corp.
1/22/2013 7:53 AM EST
Vision algorithm development
It becomes evident from the above that the best solutions will be low power, cost-optimized and high performance, both in terms of managed high bandwidth data streams and sophisticated algorithmic processing itself. However, in most cases algorithm development for ADAS applications is initially developed on a PC and often based on non-optimized algorithms found in OpenCV or provided in Matlab.
These algorithms and ADAS applications need to be ported from the initial development environment to an embedded architecture optimized for the small, low power FF ADAS processing camera system. An increasing amount of this algorithmic benchmark code is written in OpenCL, so there is a further requirement to allow reuse of this code for the embedded target. At a high level, the porting process follows the flow illustrated in Figure 1 below.
During the algorithm development phase in the PC environment, it is important to be mindful of the embedded target architecture in order to ease follow-on optimization and porting. If the embedded target software development kit (SDK) also supports a PC platform version, it is preferable to use these data structures and API whenever possible to ensure a smoother porting of the algorithm.
Once the PC algorithm is validated and performs satisfactorily, it is optimized based on the properties of the embedded platform including often a conversion from floating point to fixed point for some if not all functions. If the optimized PC algorithm adheres to the resource and power limitations of the embedded platform, the developer proceeds to port the algorithm to the target device. If not, a minimized or ‘scaled-down’ variant of the algorithm is developed and ported to the chip. The minimized variant uses approximations, simplifying assumptions and substitutes algorithmic functions with ‘lighter’ versions to produce a vision algorithm that matches the performance of the original PC algorithm as closely as possible. The optimized algorithm is then checked against an accepted test database to ensure performance benchmarks are maintained.
Tier 1 suppliers and OEMs are at different phases with respect to this development and porting process, depending on the ADAS application and specific development team. Those who have already implemented an ADAS application on a specific embedded architecture may consider alternative solutions to reduce size, power and cost. For example, the Euro NCAP requirements will drive requirements for these systems in lower-end car models. This means the existing systems potentially need to be reconsidered from the ground up.
Alternatively, some Tier 1’s and OEMs are just beginning the process of in-house ADAS algorithm development in a strategic move away from off-the-shelf alternatives. In these cases, they are at a position where, as described above, they have developed algorithms on a PC or GPU and have yet to make the leap to the embedded vision platform. In any case, component suppliers must pay special attention to easing this porting process as much as possible. Open CL support with an embedded tool chain will likely be an important ingredient to success in this application space.
Embedded Vision Alliance
The Embedded Vision Alliance is a worldwide partnership of technology providers working together to provide engineers with practical education, information, and insights to help them incorporate embedded vision capabilities into products. The Alliance’s website, at www.Embedded-Vision.com, offers a plethora of free resources including tutorial articles, videos, code downloads and a discussion forum staffed by technology experts. ADAS is key area of interest for the EVA.
Jeff Bier of the EVA has noted that with improvements in sensors, processors, algorithms and tools, it has become increasingly practical to incorporate visual intelligence into many kinds of systems, including systems with severe cost, size and power constraints, like those described here for ADAS. Because of the potential to reduce accidents and save lives, automotive safety is an especially compelling application for vision technology; hence automobile manufacturers are on the leading edge of deploying vision technology in high-volume products. However, many of the vision functions being incorporated or considered for automobiles – such as collision avoidance, augmented reality, gesture user interfaces and gaze tracking – are applicable in many other types of systems, such as robotics, consumer electronics, education and health care.
Embedded vision technology has the potential to enable a wide range of electronic products that are more intelligent and responsive than before, and thus more valuable to users. It can add helpful features to existing products. And it can provide significant new markets for hardware, software and semiconductor manufacturers. The Embedded Vision Alliance, a worldwide organization of technology developers and providers, is working to empower engineers to transform this potential into reality in a rapid and efficient manner.
About the author:
Tom Wilson is vice president of business development at Cognivue Corp., which makes image cognition processors and software for embedded vision systems. CogniVue is a founding member of the Embedded Vision Alliance.
It becomes evident from the above that the best solutions will be low power, cost-optimized and high performance, both in terms of managed high bandwidth data streams and sophisticated algorithmic processing itself. However, in most cases algorithm development for ADAS applications is initially developed on a PC and often based on non-optimized algorithms found in OpenCV or provided in Matlab.
These algorithms and ADAS applications need to be ported from the initial development environment to an embedded architecture optimized for the small, low power FF ADAS processing camera system. An increasing amount of this algorithmic benchmark code is written in OpenCL, so there is a further requirement to allow reuse of this code for the embedded target. At a high level, the porting process follows the flow illustrated in Figure 1 below.
Figure 1: Flow to Port Benchmark Algorithms
to Embedded Target
During the algorithm development phase in the PC environment, it is important to be mindful of the embedded target architecture in order to ease follow-on optimization and porting. If the embedded target software development kit (SDK) also supports a PC platform version, it is preferable to use these data structures and API whenever possible to ensure a smoother porting of the algorithm.
Once the PC algorithm is validated and performs satisfactorily, it is optimized based on the properties of the embedded platform including often a conversion from floating point to fixed point for some if not all functions. If the optimized PC algorithm adheres to the resource and power limitations of the embedded platform, the developer proceeds to port the algorithm to the target device. If not, a minimized or ‘scaled-down’ variant of the algorithm is developed and ported to the chip. The minimized variant uses approximations, simplifying assumptions and substitutes algorithmic functions with ‘lighter’ versions to produce a vision algorithm that matches the performance of the original PC algorithm as closely as possible. The optimized algorithm is then checked against an accepted test database to ensure performance benchmarks are maintained.
Tier 1 suppliers and OEMs are at different phases with respect to this development and porting process, depending on the ADAS application and specific development team. Those who have already implemented an ADAS application on a specific embedded architecture may consider alternative solutions to reduce size, power and cost. For example, the Euro NCAP requirements will drive requirements for these systems in lower-end car models. This means the existing systems potentially need to be reconsidered from the ground up.
Alternatively, some Tier 1’s and OEMs are just beginning the process of in-house ADAS algorithm development in a strategic move away from off-the-shelf alternatives. In these cases, they are at a position where, as described above, they have developed algorithms on a PC or GPU and have yet to make the leap to the embedded vision platform. In any case, component suppliers must pay special attention to easing this porting process as much as possible. Open CL support with an embedded tool chain will likely be an important ingredient to success in this application space.
Embedded Vision Alliance
The Embedded Vision Alliance is a worldwide partnership of technology providers working together to provide engineers with practical education, information, and insights to help them incorporate embedded vision capabilities into products. The Alliance’s website, at www.Embedded-Vision.com, offers a plethora of free resources including tutorial articles, videos, code downloads and a discussion forum staffed by technology experts. ADAS is key area of interest for the EVA.
Jeff Bier of the EVA has noted that with improvements in sensors, processors, algorithms and tools, it has become increasingly practical to incorporate visual intelligence into many kinds of systems, including systems with severe cost, size and power constraints, like those described here for ADAS. Because of the potential to reduce accidents and save lives, automotive safety is an especially compelling application for vision technology; hence automobile manufacturers are on the leading edge of deploying vision technology in high-volume products. However, many of the vision functions being incorporated or considered for automobiles – such as collision avoidance, augmented reality, gesture user interfaces and gaze tracking – are applicable in many other types of systems, such as robotics, consumer electronics, education and health care.
Embedded vision technology has the potential to enable a wide range of electronic products that are more intelligent and responsive than before, and thus more valuable to users. It can add helpful features to existing products. And it can provide significant new markets for hardware, software and semiconductor manufacturers. The Embedded Vision Alliance, a worldwide organization of technology developers and providers, is working to empower engineers to transform this potential into reality in a rapid and efficient manner.
About the author:
Tom Wilson is vice president of business development at Cognivue Corp., which makes image cognition processors and software for embedded vision systems. CogniVue is a founding member of the Embedded Vision Alliance.
|
|
|
|
|
Navigate to related information
William Miller
3/19/2013 11:24 AM EDT
From one hand, this driver-assistance will help in different situations. From the other hand, year by year a driver is less and less capable to react on the road. Machines take leading positions, soft rules the speed and braking. I'm afraid In 10 years all cars will be autonomous. I say it's dangerous!
______________
William - http://www.carid.com/
Sign in to Reply