Automotive OEMs, their supplier partners and governments around the world have been working diligently to develop and promote active safety and advanced driver assistance systems (Adas) designed to either increase accident avoidance or reduce the severity of automobile crashes.

According to the World Health Organization, the number of automotive-related deaths worldwide in 1998 was 1.2 million. In 1990, automotive-related accidents were the ninth-leading cause of death worldwide, and they are projected to continue to increase, becoming the third-leading cause of death by 2020.

Touted by automotive analysts as the top new technology for 2010, Adas increases driver awareness of possible hazards, potentially improving reaction times with such features as lane departure warning systems, drowsiness detection and night vision. As consumers become educated about the greater safety provided by Adas, high consumer acceptance is expected to drive market demand for these features. Adas has already begun to be implemented in luxury cars, and as the technology matures and trickles down to mass-market vehicles, higher volumes will bring great cost economies. For automotive OEMs, active safety and Adas provide the added benefit of offering important product differentiation, given that today's passive safety systems are quickly becoming standard features.

Adas at a glance
Adas are not designed to take control of a vehicle. Rather, the systems increase safety by providing the driver with relevant information about the environment and operating conditions surrounding a vehicle and by alerting drivers to hazards before the drivers might notice them.

Adas applications use a variety of sensors to collect physical data about the vehicle and its surroundings. After collecting this data, an Adas system will employ object detection, recognition and tracking processing techniques to evaluate threats. Two example applications are lane departure warning systems, which alert the driver when an unintentional lane change is detected, and traffic sign recognition. When implementing lane departure warning systems, the system detects and tracks road lanes relative to vehicle position to notify an inattentive driver that the vehicle is crossing over into the adjacent lane. For traffic sign recognition, the system reads traffic signs to provide drivers with the last speed limit or notify them if they are traveling in a particular zone.

Systems usually require the use of different sensor types to collect environmental information. Lane departure warning uses CMOS camera sensors, night vision uses an infrared sensor, adaptive cruise control typically utilizes radar, and ultrasound aids in parking assist. Although the details of each application vary, the processing usually consists of three stages: data capture, preprocessing and post-processing. Preprocessing involves functions that apply to the full image and are therefore data-intensive and regular in structure. These include transformations of the image, stabilization, feature and signal enhancements, noise reduction, color conversion, motion analysis and many more. The post-processing stage involves feature tracking, scene interpretation, system control and decision making.

The process of recognizing, tracking and evaluating driving-related objects is a complex one. Driving styles and conditions affect the quality of raw data collected by sensors and can obscure important details necessary for recognizing and tracking objects. Drivers operate their vehicles in a highly dynamic and unpredictable fashion under a variety of weather conditions, including bright sunlight, rain, fog and snow. To complicate matters, all processing must be done in real-time with processing latency no greater than 30 milliseconds.

Different sensor systems may be deployed in a vehicle to perform different functions.

Each step from data capture to action requires substantial signal-processing capabilities, making high performance essential for implementing active safety and Adas accurately and in a timely fashion. Digital signal processors (DSPs) specifically designed and optimized for automotive safety applications provide the needed performance, enabling OEMs to bring active safety and Adas to market.

Dynamic flexibility
In addition to high performance, Adas applications require a flexible architecture. For example, traffic signs vary from country to country by language, text font, shape and color. Flexibility is also needed to maximize reuse of intellectual property across product lines and to cost-effectively manage the quick pace of innovation typical in any emerging market.

Innovation is most efficiently captured in software, and software-programmable architectures in processors give developers the flexibility necessary to support changing algorithms.

Consider the preprocessing algorithms used to filter out the effects of driving conditions, such as bright sunlight. Some automotive suppliers use a single algorithm; others use one algorithm for day processing and another for night driving. In reality, a wide range of pre- and post-processing algorithms may be necessary to cover the large variety of driving conditions. The system must also be able to adapt quickly, such as when a vehicle enters a tunnel, effectively changing from day driving to night driving in an instant.

Note that multiple sensors around a vehicle must perform different functions. Laterally facing sensors handle blind-spot detection; front-facing sensors manage vehicle, lane, traffic sign and pe- destrian recognition; and sensors inside the vehicle perform occupancy sensing and detect driver drowsiness and intent.

Additionally, different sensors process different kinds of data. Some traffic sign recognition algorithms rely on sign color. In those cases, forward-facing sensors need to support a wide color scale. On the other hand, grayscale sensors are much more sensitive to variations in brightness, and they offer nearly double the spatial resolution of a color sensor. Most Adas functions rely on high sensor sensitivity, so a grayscale camera is a better fit. It is also important to note that imaging sensors for Adas applications usually have a high dynamic range, normally exceeding 8 bits per pixel.

The most efficient way to handle these challenges is to execute multiple algorithms on a single DSP. For example, a forward-facing image sensor can provide the video information necessary to implement both lane departure warnings and traffic sign recognition.

Ideally, a single DSP can perform all driving condition preprocessing, as well as handle recognition tasks such as lane departure warning and traffic sign recognition. This reduces chip count, leading to fewer points of failure, increased system reliability and lower system cost--all key drivers for automotive applications.

The most robust Adas implementations will also coordinate all of the active safety subsystems in a vehicle. For example, where a driver is facing and focused has a direct impact on the effectiveness of traffic sign alerts. If a system warns of an upcoming stop sign prematurely--before the driver has a chance to recognize the sign and begin to slow down--it will be more of a nuisance than a helpful driving aid. Thus, before a warning alert is issued, information should be gathered from the traffic sign recognition system and from the interior driver-monitoring system. When the driver is facing in the direction of the road--as evaluated by the driver-monitoring system--the stop sign alert does not need to be triggered as quickly.

A robust Adas system can even evaluate complex driving circumstances. For example, if the vehicle is quickly app-roaching a stopped or slowed vehicle, a fast lane change is probably necessary. In this case, suppressing the lane departure alert will avoid distracting the driver from completing the maneuver. Of course, if a blind spot monitoring system detects another vehicle beside the car, the system should override the suppression.

SoC design efficiencies
Today's system-on-chip (SoC) architectures enable further efficiencies by integrating all of the peripherals necessary for a complete video/imaging-processing system within a single chip. With a wide range of peripheral support, today's highly integrated devices also make it easy to connect to the rest of the vehicle's systems. For example, systems-on-chip can provide the direct video output required for applications such as rear-view parking assist. They can also provide direct connection to the vehicle's main control system through an appropriate bus such as CAN, LIN or FlexRay.

System-on-chip architectures also provide application-specific specialization without the cost of an ASIC implementation. Done right, an SoC can also maintain the flexibility of a programmable software architecture, as opposed to the rigidness of an ASIC.

For example, TI's DaVinci processors include a powerful video front end that offloads key preprocessing functionality from the main CPU. Specifically, the video front end offers a resizer block that can resample (upscale or downscale) images to the appropriate resolution without consuming DSP cycles. Resizing is needed because the size of an object relative to the video frame changes as the vehicle approaches the object.

The DaVinci video front end also offers a histogram function to provide a pixel intensity distribution for each video frame. The pixel intensity distribution provides feedback about the quality of the captured image. For example, if the image is too dark, the DSP can adjust the contrast to improve processing accuracy. The front end can also take care of color space conversions without involving the main CPU. Together, these integrated blocks offload the CPU, enabling developers to implement more value-added Adas functions on a single DSP.

The SoC architecture should also be designed to move data efficiently. As with any video application, the more often data must be moved, the greater the latency of processing. In order to increase system performance and maximize the use of level-one memory resources, developers usually limit processing to areas of interest. Focusing processing on specific areas of interest allows the image block that needs to be processed be significantly smaller than when the entire image is processed and evaluated. For example, when road lanes are identified and tracked, the sky above the road does not contain pertinent data, and therefore that section of the frame is discarded.

To support this type of data movement, the SoC needs a multichannel, multithreaded direct memory access engine. The DMA controller should support a wide variety of transfer geometries and transfer sequences. Transfers on earlier DMA controllers (such as the EDMA2 controller on previous TI chips) were limited to only two dimensions, and they shared the same index parameters for source and destination. In contrast, the EDMA3 controller on the DaVinci processors supports independent source and destination indexes, as well as three-dimensional transfers.

In addition to video input and processing, applications such as parking assistance also require video output. Video output capability can also be very useful during research and development and system debug stages, even if video-out is not planned for production. To support video output, the DaVinci processor includes a video-processing back end comprising an on-screen display engine (OSD) and a video encoder. The OSD engine is capable of handling two separate video windows and two separate OSD windows. The encoder provides four analog video outputs and up to 24 bits of digital output in various formats.

Adas applications are cutting-edge, rapidly evolving technology, so developers need tools that simplify development and aid rapid prototyping. Thus, Adas algorithm development works best when using C or modeling software such as Simulink or Matlab. Of course, a working system needs more than just algorithms., so it is essential to have off-the-shelf software components such as the real-time kernel and peripheral drivers. It is also helpful to have off-the-shelf application-specific development tools and algorithm libraries. Those components can cut Adas development time by months and will be supported on DaVinci processors.

Quality standard
Finally, suppliers that want to place their products within automotive applications need to achieve AEC-Q100 qualification, the industry quality standard for integrated circuits. It is extremely difficult to achieve this qualification unless the solution was built with this specification in mind. By selecting components that are designed to be AEC-Q100-compliant, suppliers can ensure that they will successfully meet automotive quality goals.

Today's engineers and business leaders are being called on to step forward and drive down the number of automotive-related deaths through active safety and Adas. Through the use of high-performance SoCs based on flexible DSP architectures, application-specific development tools and libraries, and innovative algorithms, automotive suppliers and OEMs can bring robust Adas applications to market.

 See related image

Brooke Williams (brookew@ti.com is automotive and machine vision marketing manager for the DSP group at Texas Instruments Inc.