GM ACAS Cars Getting Ready for the Road (01/03)

The GM/NHTSA Automotive Collision Avoidance System (ACAS) project is one of the most ambitious activities within the US Department of Transportation's Intelligent Vehicle Initiative program -- $35M of federal and private funds to evaluate the safety performance and driver interface of a combined Adaptive Cruise Control (ACC) and Forward Collision Warning (FCW) system implemented on 2002 Buick LeSabre automobiles. Engineering development to date has focused on algorithms, sensors, sensor fusion, a data collection subsystem, test methodologies, and pre-production-level engineering integration of the equipment into the LeSabres. A fleet of these vehicles will be handed over to individual drivers for regular usage over a period of weeks, to assess both the human interface and technical performance.

ACAS is a joint partnership of General Motors, Delphi Automotive Systems, and the National Highway Traffic Safety Administration (NHTSA). The Volpe National Transportation Systems Center is serving as independent evaluator. Veridian and the University of Michigan Transportation Research Institute (UMTRI) provide support under contract to GM.

In an exclusive interview last month, IVsource talked to Jack Ference, ACAS federal project manager within the National Highway Traffic Safety Administration (NHTSA) and Dr. Jeremy Salinger of Veridian, Technical Director of the ACAS effort on behalf of GM and its partners. The conversation covered recent activities and next steps.

The ACAS system approach can be considered a "second-generation" collision avoidance system, as it integrates 76 GHz radar for obstacle detection, machine vision for lane detection, and Differential GPS satellite positioning using a commercial map database. "First generation" systems fielded in the nineties by other industry leaders (such as Eaton VORAD) have relied on radar alone.

Delphi has been primarily responsible for development and integration of the braking system, radar, vision, ACC, heads-up display, and path prediction. GM engineering has had responsibility for threat assessment, data fusion, digital map integration, and system integration.

One unique aspect of this project is the incorporation of a collision warning algorithm that was developed by the Johns Hopkins University Applied Physics Laboratory under NHTSA funding. This is the only known case of collision warning algorithm development funded 100% by a government agency. Why was the APL algorithm developed? “The government-developed algorithm allows us to better estimate benefits of the system. We can ‘get inside it’ in a way that is not possible with the proprietary GM algorithm” says Mr. Ference. “The algorithm will run in parallel with the operating GM algorithm and generate mock alerts into the database. In this way, we have an excellent opportunity to compare the two algorithms.”

Dr. Salinger provided a review of progress since the last IVsource update in mid-2002. During June-August, the project partners completed and tested the pilot vehicles, which incorporate the final technical design with near-final packaging.

Work then moved forward in two tracks: producing deployment vehicles and human factors (HF) testing.

For the HF testing, normal drivers drove the vehicles, accompanied by researcher observers. The purpose was to assess system and driver behavior, both objectively and subjectively. Final stage HF testing is currently underway, with an emphasis on ensuring that the data collection suite is working properly. According to Dr. Salinger, drivers have generally found the system “easy to understand” while driving the vehicle over a period of several hours.

So far in the vehicle production process, six vehicles have been completed out of a total of 13 planned. All vehicles are expected to be completed by the end of January. The production LeSabre vehicle subsystems are being used, with the exception of automatic braking provided by auxiliary parallel brakes.

The design of the field operational test (FOT) calls for a start in February or March and a wrap-up nine months later. Drivers selected from the public will have unlimited usage of the vehicles for six weeks each, so that over 70 drivers will act as test subjects. It is expected that total mileage logged will be upwards of 50,000 miles.

Basic data will be collected continuously by the in-vehicle data collection subsystem. This data is downloaded by project personnel at the end of the driver’s six week run. Additionally, a small amount of data is sent wirelessly with each collision warning alert.

When alerts occur, video before and after the event is captured. Video cameras look at both the driver and the forward scene (using a dedicated camera with wide field-of-view).

Noting that “no system is perfect,” Dr. Salinger described the team’s approach to minimizing false alarms. Generally, he says, false alarms relate to either uncertainty about the road path or driver intentions. Substantial progress has been achieved in reducing false alarms – a reduction of 50% compared to the performance of the system one year ago. “Improvements have been achieved mainly from better understanding of driver intentions and a better rule set [algorithm],” says Salinger.

The mix of sensors provides a strong “defense” against false alarms. In particular, machine vision and digital mapping / positioning are complementary -- the vision subsystem can easily detect mild curves in the road, while the map / positioning subsystem can provide information on sharp curves not easily discernible by the vision system. The system also tracks traffic ahead of the vehicle to estimate the vehicle path, using the Delphi radar and patented software.

The team has defined a standard route in the Detroit area which includes all road types. This route is used in each system evaluation of false alarms, so as to have a standard baseline. An approximate (and unofficial) goal is to experience no more than one false alarm per hour, which is considered quite challenging for a system of this type. Dr. Salinger pointed out that, to the customer, acceptability of false alarms depends on context – random false alarms are likely to be more disturbing than predictable ones (such as the same spot on the road every day during the morning commute).

The false alarm performance requirements have affected system functionality, as well. The system does not brake for stopped objects (such as a stopped vehicle in the road), although the sensors are capable of detecting them. This it because it is considered too hard to project the vehicle path relative to the object and therefore could create too many false braking situations. False braking is the worst type of false alarm, as drivers and passengers are startled and alarmed, and the vehicle can even by rear-ended by a following vehicle in the extreme case.

A key research question is whether drivers will be able to understand and distinguish these subtleties in system performance.

What is the government trying to accomplish in this project? “We want to understand what impacts ACC and FCW have on driving,” says Mr. Ference. He noted that current ACC systems have braking, whereas ACC systems evaluated in the nineties by NHTSA did not have braking. Thus, evaluating differences in ACC performance and usefulness will be of interest to the researchers.

Mr. Ference said an interim project report should be out by April 2003. A compact disk containing reports to date published by the project will be available at the Federal Highway Administration booth at the Transportation Research Board annual meeting in January.

NHTSA has just published a report on the government-developed algorithm described above. It is entitled “NHTSA Read-End Collision Warning Algorithm – Final Report.” The document number is DOT HS 809 526, SEPTEMBER 2002. The report synopsis is as follows:

As part of its ongoing research activities supporting the development, testing and evaluation of rear-end collision warning systems, NHTSA developed an experimentally-based rear-end collision warning algorithm and sponsored analysis of the algorithm using test data collected during an intelligent cruise control field operational test. In October of 1999, NHTSA sponsored a joint research effort with The Johns Hopkins University Applied Physics Laboratory to refine the algorithm previously developed by NHTSA and to translate it into structured C code capable of running in the real-time processor environment of a prototype vehicle-based system. This report describes the theory of operation of the NHTSA Algorithm and its theoretical performance under a defined set of typical rear-end collision scenarios. Test data from a series of controlled test track studies is discussed and an estimate of the number of rear-end collisions that could be avoided using the NHTSA Algorithm is also presented.

Copies of this report may be downloaded from NHTSA R&D's website at: www-nrd.nhtsa.dot.gov/departments/nrd-12/pubs_rev.html

Alternatively, a copy of the report in CD or printed format can be requested from Mr. Ference (Jack.Ference@nhtsa.dot.gov).

Alternatively, a copy of the report in CD or printed format can be requested from Mr. Ference at NHTSA (Jack.Ference@nhtsa.dot.gov).