[论文] AI applications preventing traffic accidents

严正声明:本篇论文是站长本人原创,并已通过境外平台发布。严禁任何形式的搬运和抄袭行为。如有任何疑问请联系我:[email protected]

注:此篇论文原文为全英文版本,点此查看中文翻译版

Chris Li
XXXXXX School Shanghai, China

Abstract:

With the development of self-driving technology in automotive industry, transportation safety has a very close connection with AI. There are two main approaches for the AI applications for traffic accidents. One is the analysis and solving of the potential risk of traffic accidents. Another is maximizing the safety of people’s lives if the accident cannot be avoided. Both of the works act as an important part to prevent traffic accidents. They require applications’ analysis and making decisions within a short time. This paper is introduced the methods for AI dealing with traffic problems, which is the most important part of applications’ processing. Optimizing the codes and decreasing the reaction time of the applications is also necessary to maximize saving people’s lives.

Keywords: recognize; process; reaction; training model; AI, traffic accidents, reaction time, detection, identification.

1. Introduction

Since the first real self-driving car was manufactured in the 1980s, scientists have been always dedicating themselves to improving the ability of self-driving systems. With the quick development of embedded computers and artificial intelligence, in-car computers now are more likely to simulate driving scenarios and predict the potential risks during driving.

Additionally, according to the SAE International[i], six levels of driving automation were defined, from Level 0 to Level 5, which are gradually from limited driver support features to automatic driving under all conditions.

Additionally, according to NHTSA[ii], about 93% of traffic accident is caused by human beings[iii]. People are likely to distract while driving, but programs seem never to make mistakes. Therefore, in theory, self-driving systems can greatly reduce the frequency of accidents.

SAE J3016 Levels of driving automation graph

2. Methodology

There are two main sources of risk to driving, environmental and human factors. Environmental factors are generated by the driving environment itself. For example, low visibility conditions due to foggy weather and icy roads make it difficult to get control of vehicles. Human factors are mainly non-compliance with traffic rules and unexpected actions from human beings. For example, irregular lane changes, not following the marked direction of travel and do not follow the instructions of the traffic lights. Under these all kinds of constructions, AI must find an appropriate way to avoid accidents happen or cause minimal casualties.

Traffic Accident on icy road[iv]

The whole system is made up of three main parts. The first part is detractors monitor the whole environment around the car. Cameras are used for the detection and identification of signs, markings on the ground, and tail light status of the vehicle in the front, etc. LIDARs are used for detecting the distance of obstacles ahead and the objects around the cars. Ultrasonic Radars are used for detecting obstacles in low-speed scenarios, such as in the process of entering the vehicle into the parking space.

The second part is the embedded system analyzes the data collected by the sensors and calculates and identifies the potential risks in the future driving process, making plans to respond to these threats.

The third part is to mobilize the vehicle’s brakes, steering wheel and engine when necessary.

3. Discussion to avoid accidents

From the traditional man-drive car gradually intelligent is a long-term process. Currently, The highest level of autonomous driving that has been marketed is L2, which still required drivers put their hands on the steering wheels and only provide help in specific scenarios[v].  There is still a long way to go before people want to use AI to avoid traffic accidents. The road environment is highly variable, so AI needs to consider many factors to avoid accidents to the maximum extent possible.

At this stage, the ai safety system plays the role of a watchdog in driving. It automatically changes the vehicle’s trajectory and reduces speed as much as possible when pedestrians or obstacles appear in front of the vehicle. However, that requires plenty of calculation before the AI does these operations. Before turning the steering wheel, AI should consider if the speed is safe to change direction. Otherwise, there is a risk of the vehicle overturning or losing control. It should use the LIDARs to detect the distance to the vehicle behind the adjacent lane to prevent danger from the rear.

autonomous-cars-lidar-sensors

[vi]

There are also extreme cases where a crash is unavoidable. At this point, AI should try to avoid injury to the greatest extent possible. AI can try to pop the airbag in advance, tighten the seat belt, and use the more protective parts of the vehicle for impact.

From the perspective of the historical development of technology, the car computer will have a great possibility to access the Internet in the form of the Internet of things. If that happened, cars can access and share real-time road information from a unified platform. Based on this information, the driver can anticipate possible accidents and risks ahead, optimize the driving route and alert the driver. In some unusual roads, AI can give some driving suggestions by analyzing the historical data of previous vehicles, such as braking early, changing lanes, watching out for cars at the entrance and exit of the road, etc. Most serious accidents are caused by the accumulation of multiple error factors, which may be avoided if any one factor is addressed in advance.

The driver’s factors are also an important cause of accidents, such as drunk driving, fatigue or extreme emotions. According to the NBSC[vii], in 2008 in China, the number of traffic accident deaths caused by drunk driving accounts for 4.16% of the total number of traffic accident deaths[viii]. Therefore, in addition to monitoring the state of the road, AI should also include the detection of the driver. In case of frequent abnormal driving conditions, such as multiple-lane departures, the driver should be advised to stop and rest. When road conditions allow and technology permits, autonomous driving services can be provided to reduce the burden and fatigue of drivers driving for long periods. When a driver is detected as a risk of driving under the influence of alcohol, drugs or substances, the driver should be discouraged and offered a viable solution by providing navigation planning for nearby rest stops or calling a chauffeur on his or her behalf. Automatically calls the driver’s family members or police for assistance when necessary.

According to the Volvo Trucks Safety Report 2017[ix], 1.2 million people are killed in road traffic accidents Worldwide each year. 60% of truck and pedestrian (or bicycle) collisions occur in urban areas. Therefore, it seems that traffic accidents are more likely to occur in cities. Although the speed in the city is slower, the driving scenario on the street is more complex and requires the driver to concentrate on various unexpected situations. In fact, there are more feasible solutions to traffic accidents in the city as well. The slower speed allows the vehicle to slow down faster, leaving the driver more time to react.

Volvo Trucks Safety Report

In modern cities, traffic management usually sets up some monitoring and speed-measuring devices. It would be useful if these devices could be connected to AI systems for traffic accident prevention and avoidance.

The first step is to interconnect the in-vehicle system with the municipal information system. For a faster and more stable connection, this system is not necessarily connected to the server through the cellular mobile network as in traditional communication, but rather to the communication system on the road facilities nearby.

The in-vehicle system then exchanges information with the street facility. The onboard system sends information such as the car’s route and speed to the street traffic facility, and the street facility sends several surrounding intersections and street information to the car, such as speed information of vehicles ahead, and pedestrian information.

This can solve the problems caused by blind spots and narrow roads in streets with complex driving conditions, and also includes pedestrians and non-motorized drivers into the consideration of the entire system, not just other motor vehicles.

Meanwhile, in order to save the computing resources of the car AI system, the above-mentioned computing part should be performed by the ground computing system as much as possible and send the calculation results directly to the car, which will then decide the next operation and send its decision to the street system again for reference by other vehicles. Also, reminder signs can be set up on the streets, such as on some narrow streets, to alert other pedestrians and non-motorized vehicles when a vehicle is passing.

Large trucks cause an important part of the accident. According to the NHTSA “Traffic Safety Facts 2014[x]”, In terms of crash rates, large trucks are higher than family buses. The fatal crash rate for buses in 2014 was 1.34 compared to 1.29 for family buses, with family buses having a lower accident rate than large trucks.

Compared with ordinary cars, trucks have more volume and mass, which makes them more difficult to maneuver, braking and turning takes longer. This is exactly why they cause more accidents. In the prevention of such accidents, we should give higher priority to these large vehicles and give more to smaller vehicles to avoid these accidents. When they can’t stop, they can send a signal to the vehicle in front of them, making the AI in front of them to avoid tailgating accidents automatically. Such messages will not be sent to a specific vehicle, but will be broadcast to all cars in the surrounding area without discrimination.

4. Discussion to minimize injuries

In the above paragraphs, some ways to avoid accidents are explained. However, in some cases, collisions and injuries are always unavoidable. At this point we should seek some options to minimize injuries.

In the current design scheme, the more common method of reducing damage is mainly through strengthening the body structure, equipped with occupant protection devices (such as airbags) to achieve this purpose. Most of these protective measures are reactive and therefore have many limitations. There will always be some special circumstances that prevent the protection from taking effect. AI can protect passengers by intervening proactively.

The impact force from the collision is the main cause of the injury. If the AI can reduce the impact, the injury can be reduced. When impacting, try to avoid bulky vehicles or hard objects, and absorb energy by hitting barriers that can provide a cushion such as guardrails and green belts. And use the angle of the vehicle with better protection to face the obstacle, avoid contacting the obstacle from the side and offset the angle. The AI can analyze material information from the vehicle ahead as well as from surrounding obstacles. Avoid collision with trucks, buses, etc. transporting hazardous materials and choose grass and woods instead.

Side Mobile Barrier test[xi]

In high-speed driving, if you detect too much speed to avoid a crash, you can first tighten the passenger seat belt, lower the windows to prevent injuries from glass shards, and issue a warning alert to alert rear passengers. Buy them a few seconds to prepare for the impact in advance and use some collision avoidance positions to avoid injury.

After an accident, the AI should proactively call the rescue service and the driver’s family members, send the vehicle’s location, and provide, via the cloud platform, the damage to the vehicle, the injuries of the members, and the health information of the accompanying members saved in advance: such as age, blood type, drug allergy information, and historical diseases. In addition, it automatically turns on the vehicle’s double flashes and automatically uses the vehicle’s horn system to draw the attention of people passing around and seek help.

5. Conclusion and Future Work

All of the above calculations require the AI to have a correct estimate of the vehicle’s state, weight, performance and other data. This usually takes a long time to test and learn before it can be used in real-life. This poses a great challenge to the learning ability of the AI, as each model has different performance conditions and different models of vehicles require targeted learning, testing and optimization.

AI can record drivers’ driving habits to improve prediction accuracy. The driver information generated by long time accumulation should be exportable and can be migrated and used synchronously on AI systems of the same architecture. Vehicle manufacturers should create a neutral organization to share information, share road information with the user’s consent, and make it easier for users to use their own personalized profiles across vehicles.

In addition to local-based computing, it is also possible to try to establish connections between vehicles using communication devices, allowing the AI to pass information and cooperate better in case of danger. During normal driving, AI in neighbor cars can use the peer-to-peer networking to communicate with each other about planned routes, so that potential risks can be detected in advance and accidents can be stopped.

Some of the ideas proposed above about AI applications to prevent traffic accidents are in the process of human practice, and some designs are still in the imagination stage. AI is the product of model training, so people still need more cases and experiences to optimize AI. At the same time, communication between vehicles requires close cooperation between vehicle manufacturers, and the system will be extremely useful when a certain percentage of vehicles can communicate on the road. In this regard, mankind still has a long way to go.


[i] Society of Automotive Engineers
[ii] National Highway Traffic Safety Administration (US)
[iii] https://one.nhtsa.gov/people/injury/research/udashortrpt/background.html
[iv] https://www.cbc.ca/news/canada/manitoba/rutted-icy-roads-crashes-winnipeg-1.6309643
[v] https://www.sae.org/news/2019/01/sae-updates-j3016-automated-driving-graphic
[vi] https://www.technologyreview.com/2017/03/20/153129/autonomous-cars-lidar-sensors/
[vii] National Bureau of Statistics of China
[viii]https://transport.ckcest.cn/CatsCategory/listGLJCSJ3?tableName=cats_highwayac_reason&pubflag=1&code=C09&pageNo=1
[ix] https://www.volvogroup.com/en/about-us/traffic-safety/most-common-accidents.html
[x] https://crashstats.nhtsa.dot.gov/Api/Public/ViewPublication/812261
[xi]https://www.euroncap.com/en/vehicle-safety/the-ratings-explained/adult-occupant-protection/lateral-impact/side-mobile-barrier/

中文翻译版

注:此版本为机翻,部分内容可能与原文不符,仅供参考。

AI在预防交通事故方面的应用

摘要:随着汽车行业自动驾驶技术的发展,交通安全与AI 有着非常密切的联系。AI 在交通事故中的应用主要有两种途径。一是对交通事故潜在风险的分析和解决。另一个是在事故无法避免的情况下,最大限度地保障人们的生命安全。这两项工作都是预防交通事故的重要组成部分。它们要求应用程序在短时间内进行分析并做出决定。介绍了应用处理中最重要的环节——交通问题的AI 处理方法。优化代码和减少应用程序的反应时间也是必要的,以最大限度地拯救人们的生命。
关键词:识别;过程;反应;训练模型;AI,交通事故,反应时间,检测,识别。

1.介绍
自上世纪80 年代第一辆真正的自动驾驶汽车制造出来以来,科学家们一直致力于提高自动驾驶系统的能力。随着嵌入式计算机和人工智能的快速发展,现在的车载计算机更倾向于模拟驾驶场景,预测驾驶过程中的潜在风险。

此外,根据SAE International (i) 定义了6 个级别的驾驶自动化,从0 级到5 级,逐步从有
限的驾驶员支持功能到所有条件下的自动驾驶。另外,根据NHTSA (ii),约93%的交通事故是由人类造成的 (iii) 。人们在开车时容易分心,但程序似乎从来不会出错。因此,理论上,自动驾驶系统可以大大降低事故发生的频率。

SAE J3016 级驾驶自动化图

2.方法
驾驶风险主要有两大来源:环境因素和人为因素。环境因素是由驾驶环境本身产生的。例如,由于大雾天气和结冰的道路导致能见度低,车辆很难得到控制。人为因素主要是不遵守交通规则和人为的意外行为。比如不规则变道、不按照标示的行驶方向行驶、不按照红绿灯指示行驶等。在这些各种建筑下,AI 必须找到适当的方法来避免事故发生或造成最小的伤亡。

结冰道路上的交通事故

整个系统由三个主要部分组成。

第一部分是诋毁者监控汽车周围的整个环境。摄像头用于检测和识别地面上的标志、标记、前方车辆的尾灯状态等。激光雷达用于探测前方障碍物和汽车周围物体的距离。超声波雷达则用于检测低速场景下的障碍物,比如在车辆进入车位的过程中。

第二部分是嵌入式系统对传感器采集的数据进行分析,并计算和识别未来驾驶过程中的潜在风险,制定应对这些威胁的计划。

第三部分是在必要时调动车辆的刹车、方向盘和发动机。

3.避免事故的讨论
从传统的人驾驶汽车逐渐智能化是一个长期的过程。目前已经推向市场的最高水平的自动驾驶是L2,它仍然需要驾驶员把手放在方向盘上,只在特定场景下提供帮助。在人们希望使用AI 来避免交通事故之前,还有很长的路要走。道路环境变化很大,因此AI需要考虑多方面的因素,以最大限度地避免事故的发生。

在这个阶段,ai 安全系统在驾驶中扮演看门狗的角色。它会自动当行人或障碍物出现在车辆前方时,会改变车辆的轨迹,并尽可能降低速度。然而,在AI 进行这些操作之前,这需要大量的计算。在转向方向盘之前,AI 应该考虑速度是否安全,以改变方向。否则,就存在车辆翻车或失控的风险。它应该使用激光雷达来检测与相邻车道后面车辆的距离,以防止来自后面的危险。

自动驾驶车辆的光学雷达

在某些极端情况下,崩溃是不可避免的。在这一点上,AI 应该尽量避免受伤。AI 可以尝试提前弹出安全气囊,收紧安全带,并使用车辆更具有保护作用的部分来应对冲击。

从技术的历史发展来看,汽车电脑以物联网的形式接入互联网,将有很大的可能性。如果这种情况发生,汽车就可以从统一的平台上获取和共享实时的道路信息。基于这些信息,驾驶员可以预测前方可能发生的事故和风险,优化驾驶路线,并向驾驶员发出警报。

在一些不寻常的道路上,AI 可以通过分析过往车辆的历史数据,给出一些驾驶建议,如提前刹车、变道、注意道路出入口的车辆等。大多数严重的事故都是由于多个误差因素的累积造成的,如果提前解决其中任何一个因素,就有可能避免这种情况。

驾驶员因素也是造成事故的重要原因,如酒后驾车、疲劳或情绪极端。据NBSC(vii) 统计,2008 年我国因酒驾导致的交通事故死亡人数占交通事故死亡总人数的4.16% (viii)。因此,除了监测道路状态外,AI 还应该包括对司机的检测。如果出现频繁的异常驾驶情况,如多车道偏离,应建议驾驶员停车休息。在路况允许、技术允许的情况下,可以提供自动驾驶服务,减轻司机长时间驾驶的负担和疲劳。当司机被检测出有在酒精、药物或物质影响下驾驶的风险时,应劝阻司机,并通过为附近的休息站点提供导航规划或为其呼叫司机提供可行的解决方案。自动拨打司机家属电话。必要时可向成员或警察求助。

根据《沃尔沃卡车安全报告2017(ix)》,有120 万人死于道路交通全球每年的交通事故。60%
的卡车和行人(或自行车)碰撞发生在城市地区。因此,似乎交通事故更容易发生在城市。虽然在城市中速度较慢,但行车场景上街道则更为复杂,需要驾驶员集中精力应对各种突发情况。事实上,城市中也有更可行的交通事故解决方案。较慢的速度可以让车辆更快地减速,让司机有更多的时间做出反应。

在现代城市中,交通管理部门通常会设置一些监控和测速装置。如果这些设备可以连接到AI 系统,以预防和避免交通事故,这将是有用的。第一步是将车载系统与市政信息系统进行互联。为了更快更稳定的链接,这个系统不一定像传统通信那样通过蜂窝移动网络与服务器相连,而是与附近道路设施上的通信系统
相连。

然后,车载系统就会与街道设施交换信息。车载系统将汽车的路线和速度等信息发送给街道交通设施,街道设施将周围几个十字路口和街道信息发送给汽车,如前方车辆的速度信息、行人信息等。

这可以解决在行车条件复杂的街道上由于盲点和狭窄道路所带来的问题,也将行人和非机动驾驶人纳入整个系统的考虑范围,而不仅仅是其他机动车。

同时,为了节省汽车AI 系统的计算资源,上述计算部分应尽可能由地面计算系统执行,并将计算结果直接发送给汽车,由汽车决定下一步操作,并再次将其决策发送给街道系统,供其他车辆参考。此外,还可以在街道上设置提醒标志,比如在一些狭窄的街道上,当有车辆通过时,提醒其他行人和非机动车辆。
大型卡车是事故的重要组成部分。根据NHTSA 的《2014 年交通安全事实》(x),就撞车率而言,大型卡车比家庭巴士要高。2014 年,公交车的致命事故率为1.34,而家庭巴士的致命事故率为1.29,家庭巴士的事
故率低于大型卡车。

与普通汽车相比,卡车的体积和质量更大,操纵起来更困难,刹车和转弯需要更长的时间。这也正是它们造成事故更多的原因。在预防这类事故时,我们应该给这些大型车辆更高的优先级,给较小的车辆更多的优先级,以避免这些事故的发生。当他们不能停车时,他们可以向前面的车辆发送信号,使前面的AI 自动避免尾随事故。这样的信息不会发送给特定的车辆,而是会无差别地广播给周围区域的所有车辆。

4.讨论如何减少伤害
在上面的段落中,解释了一些避免事故的方法。然而,在某些情况下,碰撞和受伤总是不可避免的。在这一点上,我们应该寻求一些方法来减少伤害。

在目前的设计方案中,比较常见的减少损伤的方法主要是通过加强车身结构,配备乘员保护装置(如安全气囊)来达到这一目的。这些防护措施大多是反应式的,因此有很多局限性。总会有一些特殊情况阻止保护措施生效。AI 可以通过主动干预来保护乘客。

碰撞产生的冲击力是造成受伤的主要原因。如果AI 可以减少影响,伤害就可以减少。撞击时,尽量避开笨重的车辆或硬物,通过撞击护栏、绿化带等可以提供缓冲的障碍物来吸收能量。并利用防护较好的车辆面对障碍物的角度,避免从侧面接触障碍物并偏移角度。AI 可以分析来自前方车辆和周围障碍的材料信息。避免与运输危险物品的卡车、公交车等碰撞,选择草地和树林代替。


在高速行驶中,如果检测到速度过大而无法避免撞车,可以先收紧乘客安全带,降低车窗以防止玻璃碎片伤害,并发出警示警报,提醒后排乘客。给他们争取几秒钟的时间,提前做好应对撞击的准备,使用一些避撞姿势,避免受伤。

事故发生后,AI 应主动呼叫救援服务和司机家属,发送车辆位置,并通过云平台提供车辆的损坏情况、成员的受伤情况,以及事先保存的随行成员的健康信息:如年龄、血型、药物过敏信息、历史疾病等。此外,它还会自动开启车辆的双闪,自动利用车辆的喇叭系统,吸引周围路过的人的注意,寻求帮助。

5.结语及未来工作
上述所有计算都需要AI 对车辆的状态、重量、性能和其他数据进行正确的估计。这通常需要很长时间进行测试和学习,然后才能用于现实生活中。这对AI 的学习能力提出了巨大的挑战,因为每个模型的性能条件不同,不同的车辆模型需要有针对性的学习、测试和优化。

AI 可以记录司机的驾驶习惯,以提高预测准确性。长时间积累产生的驱动程序信息应具有可导出性,并可在相同架构的AI 系统上同步迁移和使用。车辆制造商应该创建一个中立的组织来共享信息,在用户同意的情况下共享道路信息,并让用户更容易跨车辆使用自己的个性化配置文件。

除了基于本地的计算外,还可以尝试使用通信设备在车辆之间建立连接,使AI 能够在危险情况下更好地传递信息和合作。在正常驾驶过程中,邻居汽车中的AI 可以使用对等网络就计划的路线相互通信,以便提前检测潜在风险,防止事故发生。

上述提出的AI 应用于预防交通事故的一些想法正在人类实践过程中,一些设计仍处于想象阶段。AI 是模型训练的产物,所以人们还需要更多的案例和经验来优化AI。同时,车辆之间的沟通需要车辆制造商之间的密切合作,当ai一定比例的车辆可以在道路上进行沟通。在这方面,人类还有很长的路要走。

[i] 汽车工程师学会
[ii] 美国国家公路交通安全管理局
[iii] https://one.nhtsa.gov/people/ injury/ research/udashort rpt/background.html
[iv] https://www.cbc.ca/news/canada/manitoba/ rutted -icy-roads-crashes -winnipeg -1.6309643
[v] https://www.sae.org/news/2019/01/sae -updates -j3016-automated -driving-graphic
[vi] https://www.technologyreview.com/2017/03/20/153129/autonomous -cars-lidar-sensors/
[vii] 中国国家统计局
[viii] https://transport.ckcest.cn/CatsCategory/listGLJCSJ3?tableName=cats_highwayac_reason&pubflag=1&code=C09&pageNo=1
[ix] https://www.volvogroup.com/en/about-us/traffic-safety/most-common-accidents.html
[x] https://crashstats.nhtsa.dot.gov/Api/Public/ViewPublication/812261
[xi] https://www.euroncap.com/en/vehicle-safety/the-ratings-explained/adult-occupant-protection/lateral-impact/side-mobile-barrier/

暂无评论

发送评论 编辑评论


				
|´・ω・)ノ
ヾ(≧∇≦*)ゝ
(☆ω☆)
(╯‵□′)╯︵┴─┴
 ̄﹃ ̄
(/ω\)
∠( ᐛ 」∠)_
(๑•̀ㅁ•́ฅ)
→_→
୧(๑•̀⌄•́๑)૭
٩(ˊᗜˋ*)و
(ノ°ο°)ノ
(´இ皿இ`)
⌇●﹏●⌇
(ฅ´ω`ฅ)
(╯°A°)╯︵○○○
φ( ̄∇ ̄o)
ヾ(´・ ・`。)ノ"
( ง ᵒ̌皿ᵒ̌)ง⁼³₌₃
(ó﹏ò。)
Σ(っ °Д °;)っ
( ,,´・ω・)ノ"(´っω・`。)
╮(╯▽╰)╭
o(*////▽////*)q
>﹏<
( ๑´•ω•) "(ㆆᴗㆆ)
😂
😀
😅
😊
🙂
🙃
😌
😍
😘
😜
😝
😏
😒
🙄
😳
😡
😔
😫
😱
😭
💩
👻
🙌
🖕
👍
👫
👬
👭
🌚
🌝
🙈
💊
😶
🙏
🍦
🍉
😣
Source: github.com/k4yt3x/flowerhd
颜文字
Emoji
小恐龙
花!
上一篇
下一篇