IMAPCAR - the solution to automotive embedded image processing
NEC Electronics has developed an SIMD processor called IMAPCAR which stands
for Integrated Memory Array Processor for CAR. This embeds 128 8-bit processing
elements and a 16-bit control processor (see block diagram). Based on a 4-way
very long instruction word (VLIW) architecture that enables up to four instructions
to be executed per clock cycle, this processor can reach up to 100 Gops (giga
operations per second) on a 100 MHz clock. The power consumption at this performance
is very low (below 2 W). All these features are must-have requirements in the
automotive market which makes this product a perfect fit.
NEC Electronics has also developed a complete toolset (SDBIMAP), language (1DC)
and development board to make the IMAP technology easy to use. The language
is called 1DC for 1 dimension C language. It is a C-based language with SIMD
extensions for loop control, variable assignment and special instructions. For
example, a condition executed in all PEs is coded mif-melse for multiple if/multiple
else and a condition executed in the control processor is coded if-else as in
basic C language. Special operations (:> and :<) are used to shift values
between processing elements from left PE and from right PE. At the end, the
way IMAPCAR is programmed is close to vector programming with PEs and sequential
programming executed by the control processor.
IMAPCAR is the only SIMD product available on the market for vision-based automotive
safety applications. The product, already in its second generation, is used
in commercial vehicles. We invite you to join us for the next generation of
vision-based safety applications.
Vision processing to enhance automotive safety
The times when the only piece of electronics in a car was the radio are long past. Today's automotive industry has enthusiastically embraced electronics to implement electronic networks, microcontroller-based engine control, high-tech navigation systems, DVD players, and much more. But big challenges remain for electronics in automotive and safety is most probably the biggest one.
Automotive safety
Car safety has become a major preoccupation for all the players in the automotive
industry; govern- ments have passed laws designed to decrease casualty numbers
and carmakers are working on new technologies and safety systems to make their
cars safe.
Road accidents are today one of the most frequent causes of accidental death
and records show that road accidents are mainly caused by human error, like
failure to stay on the road, excessive speed, alcohol, drowsiness or lack of
attention. New safety systems are being introduced into cars to counteract these
factors. Safety systems fall into four categories (see diagram on top of the
page).
Driver assistance is used to warn the driver about car situations and help his/her
driving. Popular examples are the lane departure warning system (drift monitoring)
or park assist.
Active safety is an extension of driver assistance with a direct life-saving
connotation. It is used to prevent a driver from causing a crash or hitting
an obstacle. The main applications are pedestrian detection and collision warnings.
Eye-tracking devices can help the system to warn the driver of obstacles he
is unaware of and also prevent him from making errors through lack of attention.
Passive safety comes into action after a crash or at the point of impact. Examples
are seatbelts, airbags, or pre-crash sensing devices. These devices are tested
in the well-known crash tests, which allow an assessment of car safety levels.
Vehicle guidance tries to avoid accidents by controlling the direction and/or
the throttle/brakes of the car. Stop and go, low speed following system, and
ACC (Automatic Cruise Control) take control of longitudinal movement. Which
system is deployed depends on driving conditions, road conditions, and whether
the car is being driven on minor or major roads.
All major players believe that such applications have the potential for the
same market penetration as the airbag. Today almost 90% of new cars have more
than one airbag and there are cars that have more than ten. Another factor that
will impact on active safety systems is legislation in the various major countries.
In efforts to reduce road deaths, new laws are coming into force to deal with
specific cases. For example, pedestrian detection will be mandatory in Europe
by 2010.
There are two approaches to pedestrian detection. A shock-sensor-based system
detects and reacts to an actual collision with a pedestrian; a camera-based
system detects pedestrians as images and warns the driver of potential danger.
Vision sensing, the technology for new safety applications
This simple example shows that the sensor is a key component in any safety
system, and that system design depends directly on the sensor technology and
sensitivity. Sensors fall into two categories, active (radar, ultrasonic) and
passive (acceleration sensor, cameras).
For applications like lane departure warning systems or traffic sign recognition
there are no real alternatives to camera-based systems. Essentially, vision
is the most flexible sensing technology for car safety. Although image sensing
cannot fully replace radar, because it is very sensitive to weather conditions,
it can be used in combination with radar to create, for example, highly reliable
ACC or pre-crash systems.
The second important topic is the algorithm study. Much research and effort
has gone into setting up the algorithms for a full set of safety applications,
with vision processing as the basic technology for most of them. A typical image
processing operation can be broken down into several steps. The first is image
preparation to enhance image quality, followed by feature extraction where regions
of interest are detected and selected. Next the regions of interest are grouped
and classified. Finally, decisions are taken about these regions of interest.
The first steps are fully parallel, meaning that the same operation is performed
several thousand times, while the last steps are purely sequential.
Vision processing technologies
A number of technologies are available that deliver the performance required
for the different processing steps. The diagram below shows a comparison between
different technologies available for image processing. This comparison is based
on similar die sizes and technologies.
The general-purpose processor offers very good flexibility but is not efficient
in terms of consumption/performance. DSP technology enhances the computational
power, but is still stretched when it comes to handling complex algorithms.
These two technologies are very efficient in executing the sequential steps
of image processing but limited in performing parallel tasks.
Some wired logic (ASIC) can be used in combination with DSPs to enhance the
performance in image preparation but the system becomes complex to design. Full
wired logic is the fastest technology because, being fully dedicated to the
application, parallel and sequential tasks can be implemented. But bottle- necks
arise in terms of design time/complexity and application flexibility, because
every function implemented has a direct silicon impact.
SIMD (Single Instruct- ion Multiple Data) is a multiprocessor approach for handling
data streams as they occur in vision processing. This technology uses a set
of processing elements (PEs) that execute the same instruction at each clock
cycle. The advantages of this architecture are parallel computation, to give
the performance required during the first steps of image processing, and flexibility
because it remains a programmable architecture, executing code in RAM and capable
of performing the sequential steps. SIMD processors were widely used in supercomputers
for research in the 80s and 90s but disappeared with the onset of the frequency
race. With the growing demand for embedded hardware with low power consumption
and low frequency combined with a high level of computational power, there are
strong arguments for the return of SIMDs.
Contact
For further information please contact our Automotive Business Unit.
