Each year, there are 37,000 vehicular fatalities in the United States. An additional 2.35 million motorists/year are injured or disabled (1). While vehicle safety has steadily continued to improve, the greatest hope for safe roadways lies in the hands of engineers and computer scientists. A recent National Highway Traffic Safety Administration study suggests that autonomous technology can reduce vehicle-related fatalities by as much as 94% (2).
However, is current autonomous technology truly that advanced? Today, autonomous vehicles still result in injury and fatality rates comparable to that of human-operated vehicles. For example, consider the common metric of fatalities per vehicle miles traveled. In 2018, there were 1.14 fatalities per 100 million miles traveled by human-operated cars (3). In contrast, Waymo, the leading autonomous car developer, has seen 36 crashes in a fleet that has traveled over 10 million miles (4). Assuming that 0.7% of all car crashes are fatal (5), Waymo’s autonomous technology has 2.52 potential fatalities per 100 million vehicle miles traveled, suggesting that there is little safety advantage to autonomous cars in their current form.
In order to understand how current autonomous technology can be improved, we must first understand how it works. The AI present in autonomous systems operates through a three step process commonly known as a Perception-Action cycle (6).
The first step in the sequence, known as Perception, is for the system to detect all obstacles in its driving environment. Most current autonomous vehicles do this by using LIDAR (Light Detection and Ranging), which sends out millions of light pulses each second in order to construct a 3D map of the vehicle’s surroundings (7). Next, the vehicle uses advanced algorithms to segment the map and classify its objects into clusters. Lastly, the vehicle classifies the clusters and transmits the information to the AI. In this step, the vehicle also uses estimation techniques to carry out a process known as Localization, which consists of quantifying the vehicle’s own motion (7).
The second step, known as Planning, is where the AI decides what it will do next. This step is the most difficult, as the randomness in the movement of objects in the vehicle’s environment poses a challenge to building the discrete mathematical structures needed for the AI’s planning. Scientists have experimented with applying various stochastic structures, such as Partially Observable Markov Decision Processes (POMDP), to these complex scenarios, but they have encountered difficulties in generalizing such structures to work in all scenarios the vehicle may encounter (7). The Planning step is also limited by computational restraints. While accounting for too many motion constraints can overburden the computer, not considering enough movement limitations in the vehicle’s space can lead to crashes or other hazardous vehicle actions (7).
Lastly, the third step, known as Control, is where the vehicle actually executes what the AI has planned. During this step, sensors in the car track the physical parameters of the vehicle’s environment and generate and track an optimal trajectory to keep the vehicle safe while also conforming to the AI’s commands (7).
The greatest room for improvement in autonomous technology lies in the “Planning” step. In order to develop Level 5 vehicle automation (no human attention required), autonomous AI’s must be able to safely react to the stochasticity of nature and human actions. If a pedestrian suddenly darts across the street or a motorcyclist rapidly changes lanes into a car’s path or a lightning strike sends tree branches and street debris flying across the road, an autonomous vehicle must be ready to anticipate and handle the situation.
Currently, the predicted “revolution” that autonomous technology would have on roadway safety is yet to be seen. However, as more research is done into deep learning and artificial intelligence, autonomous vehicle technology has the potential to improve, which could significantly reduce motorist fatalities.
References:
- (2018). Road Safety Facts. Retrieved from https://www.asirt.org/safe-travel/road-safety- facts/
- Gupton, Nancy. The Science of Self Driving Cars. Retrieved from https://www.fi.edu/ science-of-selfdriving-cars
- Shepardson, David (2019, June 17). U.S. pedestrian, bicyclist deaths rise in 2018: report. Retrieved from https://wtvbam.com/news/articles/2019/jun/17/us-traffic-deaths-fall-1-in-2018 -preliminary-report/
- Rapier, Graham (2018, November 30). GM's Cruise has had the most self-driving crashes in California — here's how the autonomous rivals stack up when it comes to safety. Retrieved from https://www.businessinsider.com/cruise-waymo-apple-which-self-driving-cars-crash- the-most-2018-11.
- What Percentage of Car Accidents are Fatal? Retrieved from https://www.anidjarlevine.com/faqs/what-percentage-of-car-accidents-are-fatal/
- Gadam, Suhasini (2018, April 19). Artificial Intelligence and Autonomous Vehicles. Retrieved from https://medium.com/datadriveninvestor/artificial-intelligence-and-a utonomous-vehicles-ae877feb6cd2
- Pendleton, S.D., Andersen, H., Du, X., Shen, X., Meghjani, M., Hong Eng, Y., Rus, D., & Ang Jr., M. H. (2017, February 17). Perception, Planning, Control, and Coordination for Autonomous Vehicles. Machines. doi:10.3390/machines5010006