 |
By 2025, most cars will have partly assisted driving and steering, |
Autonomous Vehicles Are Great at Driving Straight - IEEE Spectrum
Overall Summary
Based on the analysis of over 37,000 #accidents involving #autonomous vehicles (#AVs) and human-driven vehicles (#HDVs), the study published in Nature found that AVs were generally less prone to accidents compared to HDVs. However, AVs significantly underperformed in certain situations.
Key findings:
1. SAE Level 4 AVs (capable of full self-driving without a human at the wheel) were roughly 36% less likely to be involved in moderate injury accidents and 90% less likely to be involved in fatal accidents compared to HDVs.
2. The risk of rear-end collisions for Level 4 AVs was about half that of HDVs, and the risk of broadside collisions was approximately one-fifth.
3. Level 4 AVs were nearly 50 times less likely to run off the road compared to HDVs.
4. AVs performed better than HDVs in #rain and #fog, likely due to their #radar and #lidar systems.
However, the study also highlighted two significant negative outcomes for Level 4 AVs:
1. AVs were over five times more likely to be involved in accidents at dawn and dusk, possibly due to issues with the data captured by visual sensor systems in challenging lighting conditions.
2. The odds of an accident during a turn were almost doubled for AVs compared to HDVs, suggesting that navigating turns is among the most demanding situations for an AV's artificial intelligence.
While the study's findings are generally favorable for AVs, the authors emphasize that more data is necessary to determine the precise cause of accidents and encourage others to assist in this effort. Additionally, some experts caution against making sweeping statements about AV safety based on the current available information.
Summary of original Ding & Abdel-Aty Article
This study compares accidents involving autonomous vehicles (AVs) and human-driven vehicles (HDVs) using a matched case-control logistic regression approach. The key findings are:
1. AVs tend to be safer than HDVs in many accident situations due to their faster reaction times, consistent sensing technologies, and ability to rapidly adjust speed and trajectory.
2. However, the likelihood of an AV accident during dawn/dusk conditions is 5.15 times greater than HDVs. AVs also have a 1.37 times higher probability of accidents while making turns compared to HDVs. This may be due to sensors having difficulty adapting to changing light and complex scenarios.
3. AVs have lower risks of rear-end accidents (0.41 times) and broadside accidents (0.19 times) relative to HDVs.
4. AV accidents are less likely when the vehicle is proceeding straight, running off the road, entering a traffic lane, or backing up compared to HDV accidents in the same situations.
5. AV accidents result in fewer moderate and fatal injuries compared to HDV accidents.
6. In rear-end accidents, 79% involve an HDV hitting an AV from behind, while only 21% involve an AV rear-ending an HDV. When an HDV hits an AV, the AV is usually in autonomous mode.
In summary, the study finds AVs are generally safer than HDVs, but lighting conditions and turning maneuvers remain challenging scenarios where AV accident risk can be higher. The matched case-control approach allows direct comparison between AV and HDV accident characteristics.
Where AVs fail
According to the matched case-control logistic regression analysis, there are two main scenarios where human-driven vehicles (HDVs) outperform autonomous vehicles (AVs) in terms of accident likelihood:
1. Dawn/Dusk conditions: The study found that the probability of an AV being involved in an accident during dawn or dusk is 5.15 times higher than that of an HDV under the same lighting conditions. This could be attributed to the limitations of AV sensors and cameras in adapting to rapidly changing light conditions. During dawn and dusk, the sun's shadows and reflections may confuse the sensors, making it difficult for them to distinguish between objects and identify potential hazards. Additionally, the fluctuating light conditions can impact the accuracy of object detection and recognition algorithms used by AVs, which can result in false positives or negatives. In contrast, human drivers can often better adapt to changing light conditions based on their experience and intuition.
2. Turning maneuvers: The analysis revealed that AVs have a 1.373 times higher probability of being involved in an accident while making turns compared to HDVs. This could be due to the lack of situational awareness in AVs when navigating complex driving scenarios like turning at intersections. AVs rely on sensors and algorithms to perceive their surroundings and make driving decisions, but these systems may not detect all obstacles and hazards, particularly in dynamic environments. Moreover, AVs are programmed to follow predefined rules and scenarios, which may not encompass every possible driving situation. The modifications of scenarios can present difficulty for AVs in perceiving and responding to them, thereby raising the risk of an accident. On the other hand, most human drivers have years of experience and can adapt to unexpected circumstances on the road, enabling them to make better decisions while turning, such as adjusting their speed and trajectory based on the actions of other vehicles and pedestrians.
These findings highlight the need for further advancements in AV technology to improve their performance in challenging lighting conditions and complex driving maneuvers. Improving the safety of AVs in these scenarios may require a holistic approach involving advanced sensors, robust algorithms, redundancy measures, and effective integration of sensor data. By focusing on these aspects, the safety of AVs can be significantly enhanced, bringing them closer to surpassing human drivers in all driving situations.
Where AVs Outperform
The matched case-control logistic regression analysis reveals several situations where autonomous vehicles (AVs) outperform human-driven vehicles (HDVs) in terms of accident likelihood:
1. Rainy weather conditions: The study found that the odds of an AV being involved in an accident during rainy weather are 0.334 times lower than those of an HDV. This can be attributed to the faster reaction times of AVs in responding to changing road conditions caused by rain. AVs can rapidly process sensor data and adjust the vehicle's speed and trajectory within milliseconds, whereas human drivers may take several seconds to react. Additionally, while rain can increase the likelihood of skidding or loss of vehicle control, AVs employ consistent and precise sensing technologies such as cameras, LiDAR, radar, and GPS to accurately detect and perceive road conditions, regardless of the weather.
2. Rear-end collisions: The analysis shows that AVs have a relatively lower risk of being involved in rear-end accidents compared to HDVs (0.410 times). This finding indicates that AVs can detect and react to potential rear-end collision situations much faster than human drivers can, thanks to their advanced sensors and software that can quickly analyze the surrounding environment and make decisions based on the data received. This allows AVs to respond to potential accidents before they occur, preventing or mitigating the severity of a rear-end collision.
3. Broadside (side-impact) collisions: The study reveals that AVs have a significantly lower probability (0.189 times) of being involved in broadside accidents compared to HDVs. Similar to rear-end collisions, AVs' superior sensing capabilities and faster reaction times enable them to better detect and avoid potential side-impact collisions.
4. Pre-accident vehicle movements: The analysis shows that most pre-accident movements made by AVs reduce the probability of accidents compared to HDVs. For example, when an AV is proceeding straight, the risk of an accident is 0.436 times lower than that of an HDV. Similarly, when an AV is running off the road, the accident risk is 0.325 times lower than that of an HDV. AVs can detect these situations and apply corrective actions, such as adjusting speed or steering angle, more quickly and accurately than human drivers.
5. Accident severity: The model indicates that AV accidents are less likely to result in moderate and fatal injuries compared to HDV accidents. This can be attributed to the advanced safety features and collision avoidance systems incorporated into AVs, which can help mitigate the severity of accidents when they do occur.
These findings highlight the potential safety benefits of AVs in various driving situations, particularly in terms of their ability to detect and respond to hazards more quickly and consistently than human drivers. However, it is essential to acknowledge that AVs still face challenges in certain scenarios, such as navigating complex urban environments and handling edge cases not covered in their programming. As AV technology continues to advance, it is expected that their performance will further improve, leading to even greater safety benefits compared to HDVs.
Autonomous vehicles (AVs) have made headlines in recent months, though often for all the wrong reasons. Cruise, Waymo, and Tesla are all under U.S. federal investigation for a variety of accidents, some of which caused serious injury or death.
A new paper published in Nature puts numbers to the problem. Its authors analyzed over 37,000 accidents involving autonomous and human-driven vehicles to gauge risk across several accident scenarios. The paper reports AVs were generally less prone to accidents than those driven by humans, but significantly underperformed humans in some situations.
“The conclusion may not be surprising given the technological context,” said Shengxuan Ding, an author on the paper. “However, challenges remain under specific conditions, necessitating advanced algorithms and sensors and updates to infrastructure to effectively support AV technology.”
The paper, authored by two researchers at the University of Central Florida, analyzed data from 2,100 accidents involving advanced driving systems (SAE Level 4) and advanced driver-assistance systems (SAE Level 2) alongside 35,113 accidents involving human-driven vehicles. The study pulled from publicly available data on human-driven vehicle accidents in the state of California and the AVOID autonomous vehicle operation incident dataset, which the authors made public last year.
While the breadth of the paper’s data is significant, the paper’s “matched case-control analysis” is what sets it apart. Autonomous and human-driven vehicles tend to encounter different roads in different conditions, which can skew accident data. The paper categorizes risks by the variables surrounding the accident, such as whether the vehicle was moving straight or turning, and the conditions of the road and weather.
Level 4 self-driving vehicles were roughly 36 percent less likely to be involved in moderate injury accidents and 90 percent less likely to be involved in a fatal accident.
SAE Level 4 self-driving vehicles (those capable of full self-driving without a human at the wheel) performed especially well by several metrics. They were roughly 36 percent less likely to be involved in moderate injury accidents and 90 percent less likely to be involved in a fatal accident. Compared to human-driven vehicles, the risk of rear-end collision was roughly halved, and the risk of a broadside collision was roughly one-fifth. Level 4 AVs were close to 50 times less likely to run off the road.
These figures look good for AVs. However, Missy Cummings, director of George Mason University’s Autonomy and Robotics Center and former safety advisor for the National Highway Traffic Safety Administration, was skeptical of the findings.
“The ground rules should be that when you analyze AV accidents, you cannot combine accidents with self-driving cars [SAE Level 4] with the accidents of Teslas [SAE Level 2],” said Cummings. She took issue with discussing them in tandem and points out these categories of vehicles operate differently—so much so that Level 4 AVs aren’t legal in every state, while Level 2 AVs are.
Mohamed Abdel-Aty, an author on the paper and director of the Smart & Safe Transportation Lab at the University of Central Florida, said that while the paper touches on both levels of autonomy, the focus was on Level 4 autonomy. “The model which is the main contribution to this research compared only level 4 to human-driven vehicles,” he said.
And while many findings were generally positive, the authors highlighted two significant negative outcomes for level 4 AVs. It found they were over five times more likely to be involved in an accident at dawn and dusk. They were relatively bad at navigating turns as well, with the odds of an accident during a turn almost doubled compared to those for human-driven vehicles.
More data required for AVs to be “reassuring”
The study’s finding of higher accident rates during turns and in unusual lighting conditions highlight two major categories of challenges facing self-driving vehicles: intelligence and data.
J. Christian Gerdes, codirector of the Center for Automotive Research at Stanford University, said turning through traffic is among the most demanding situations for an AV’s artificial intelligence. “That decision is based a lot on the actions of other road users around you, and you’re going to make the choice based on what you predict.”
Cummings agreed with Gerdes. “Any time uncertainty increases [for an AV], you’re going to see an increased risk of accident. Just by the fact you’re turning, that increases uncertainty, and increases risk.”
AVs’ dramatically higher risk of accidents at dawn and dusk, on the other hand, points towards issues with the data captured by a vehicle’s sensors. Most AVs use a combination of radar and visual sensor systems, and the latter is prone to error in difficult lighting.
It’s not all bad news for sensors, though. Level 4 AVs were drastically better in rain and fog, which suggests that the presence of radar and lidar systems gives AVs an advantage in weather conditions that reduce visibility. Gerdes also said AVs, unlike humans, don’t tire or become distracted when driving through weather that requires more vigilance.
While the paper found AVs have a lower risk of accident overall, that doesn’t mean they’ve passed the checkered flag. Gerdes said poor performance in specific scenarios is meaningful and should rightfully make human passengers uncomfortable.
“It’s hard to make the argument that [AVs] are so much safer driving straight, but if [they] get into other situations, they don’t do as well. People will not find that reassuring,” said Gerdes.
The relative lack of data for Level 4 systems is another barrier. Level 4 AVs make up a tiny fraction of all vehicles on the road and only operate in specific areas. AVs are also packed with sensors and driven by an AI system that may make decisions for a variety of reasons that remain opaque in accident data.
While the paper accounts for the low total number of accidents in its statistical analysis, the authors acknowledge more data is necessary to determine the precise cause of accidents, and hope their findings will encourage others to assist. “I believe one of the benefits of this study is to draw the attention of authorities to the need for better data,” said Ding.
On that, Cummings agreed. “We do not have enough information to make sweeping statements,” she said.
- The transition to autonomous-ready cars has accelerated in recent years and will continue to do so, according to Statista's Mobility Market Insights.
- There are six different automation levels for passenger cars in the classification, as this chart shows.
- By 2025, 63.6% of registered passenger cars worldwide should be at level 1, partly assisted driving and steering, Statista predicts.
The share of newly registered passenger cars worldwide produced without provisions for assisted driving systems has become smaller and smaller in recent years. According to the Statista Mobility Market Insights, “regular” cars made up a minority of only 14.4 percent of newly registered cars around the globe in 2020. The transition to autonomous-ready cars has been a quick one: In 2015, cars without assistance systems were still in the majority at 51.4 percent of all newly registered ones.
There are six different automation levels for passenger cars in the classification SAE J3016 by standards developer SAE International, four of which are included in the chart. Autonomous driving ability is displayed as levels 0 to 5. In levels 1 and 2, the autonomous driving functions assist the drivers (so-called assisted mode). Automated driving provisions of of level 3 are now entering regular production, while level 4 is expected by 2025.
Level 3 is the lowest level of this automated mode. Vehicles whose systems meet level 3 requirements can drive independently to some degree, with the driver taking over the wheel again upon request and with advance notice. From level 4, vehicles can drive independently without the driver having to take over. In levels 1 and 2, the driver can be assisted by the autonomous driving systems, yet he or she can not turn attention away from the road. In level 1, either braking or steering can be assisted. In level 2, a combination of both is possible, equipping the car lane centering and adaptive cruise control capabilities.
Discover
How is the World Economic Forum promoting sustainable and inclusive mobility systems?
Have you read?
Accept our marketing cookies to access this content.
These cookies are currently disabled in your browser.
Don't miss any update on this topic
Create a free account and access your personalized content collection with our latest publications and analyses.
License and Republishing
World Economic Forum articles may be republished in accordance with the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International Public License, and in accordance with our Terms of Use.
The views expressed in this article are those of the author alone and not the World Economic Forum.
Share:
Global Agenda
The Agenda Weekly
A weekly update of the most important issues driving the global agenda
No comments:
Post a Comment