Self Driving Car Safety Statistics Support Early Implementation

When I think about self-driving cars, the first question that pops into my mind is, “When are they going to get here?” I’m tired of driving myself around town and getting stuck in traffic! You might ask the same question if you are a software developer for autonomous vehicles or cars with advanced driver assistance systems (ADAS). Up until now, the answer has been, “When they’re extremely safe.”

In the United States, the National Highway Traffic Safety Administration estimates that vehicles with high levels of autonomy will not arrive until after 2025. This is primarily because ADAS-enabled vehicles are held to a higher standard of safety than human drivers, and this standard is extremely difficult to meet. The RAND Corporation recently released research that shows that allowing autonomous vehicles on the roads before they’re even 90% safe could help save lives. This research, and laws being considered by the US Congress, could help put self-driving cars on the roads before they meet stringent safety requirements.

Most ADAS vehicle car wrecks are caused by other human drivers.

Safety Difficulties in ADAS Enabled Vehicles

There are two main components to an ADAS-enabled car that make it difficult to guarantee safety: complex software and the possibility of hardware failures.

You know how tricky things can be on the software side because you work on it. A microcontroller is either parsing object data or interpreting raw data from a variety of sensors around the car. Programs also run all of the infotainment and other high tech features of the car. Meanwhile, your system makes decisions based on incoming information from your sensor arrays and activates safety-critical features like automatic braking. To even analyze and certify a system that is this complex you need special tools like a static analyzer. With all of these moving pieces, it can be very difficult to guarantee the safety of a self-driving car. You can try to use other tools, such as a deep neural network (DNN), to do that. However, even artificial intelligence (AI) has its faults, and it’s essentially impossible to predict with 100% certainty what it will do. To ensure safety, you can add supervisory features to check your DNN, but that adds more complexity and cost.

While bugs in your code can cause serious problems, cybersecurity can also be a safety issue. Cars with ADAS features have very large attack surfaces, which are not currently well protected. Securing embedded systems in the field is difficult even when they’re not interacting with users. Self-driving cars will certainly be used as taxis, giving passengers access to their systems and increasing the likelihood of attempted breaches.

We also need to worry about redundant systems in the hardware. Many smart vehicles now combine an array of different kinds of sensors to reduce the risk of the car incorrectly sensing its environment. We all know that car parts break all the time, which means that the sensors that an autonomous vehicle uses to navigate will eventually fail. To guarantee safety, it is important to constantly check components and have a redundant system that is ready to take over if a component fails. Many of these sensors, like LIDAR, are so expensive that manufacturers might be reluctant to provide backups. Another option is to use a parallel system like GPS and high quality maps to navigate. However, satellite positioning systems aren’t sophisticated enough to do that yet. The bottom line is that it is difficult and expensive to ensure redundancy in self-driving cars.

Fleets of cars equipped with ADAS will enable developers to improve their systems more rapidly