By Indraneel Ghosh
Earlier this week, Ubers autonomous vehicles were banned in Arizona, USA. This was imposed following an accident leading to the death of a pedestrian. At around 10 PM on March 18, one of Ubers autonomous vehicles struck a woman as she crossed a road in Tempe, Arizona. She later died from her injuries at a local hospital.
Can computer vision be still trusted?
A video released by Tempe police showed the horrifying scenes before the accident occurred. In the video, the automated Uber car was moving at a constant speed with no attempt to slow down or swerve in the moments before the collision.
Immediately following the crash, Uber suspended all autonomous vehicle testing activities across the US. But Uber is one of the many companies who is working towards designing self-driven cars. In fact, there are companies all around the world working on designing such cars. After the event, all these companies came out to say that their vehicles do not have such issues. But after witnessing the events that unfolded in Arizona, can you trust them?
This is the first major event that has raised questions about the viability of using Computer Vision and automation in its current state to address real-life problem scenarios.
Self-driving car design
Autonomous cars, or self-driving cars, as they are commonly called, work on the principles of object detection. It is a subject in the field of computer vision. These cars have mechanisms to detect objects in the surrounding and then manoeuvre around these obstacles. To further train these vehicles to drive themselves, several real-life drivers and their behaviours behind the wheel are sampled. They are then used to create a standardised model which can be used by these cars to navigate through the road. One can even say that these cars have a mind of their own. They use human behaviour combined with the ability to detect objects to travel across the roads.
What went wrong?
In case of Uber, one of the things that may have been the primary cause of the accident was that the training set used for the car was not exhaustive enough. The vehicle could not detect the presence of an object (in this case a pedestrian) and respond appropriately. Theoretically speaking, a well-designed self-driving car would have detected the pedestrian and slowed down or adjusted ensure. But as it is well known, there is quite a bit of difference between theory and practical application.
Explaining computer vision
Computer vision is one of the hottest fields of research in computer science with several research papers coming out every year. All these research papers are trying to figure out newer and more efficient methods for handling the problem. Companies around the world are attempting to use it to get rid of the errors that are caused by humans. This incident seems to suggest that we are still quite far from achieving that goal. But, it also offers an opportunity to develop more efficient object detection techniques. It is an example of what might go wrong. It allows the companies to address this problem in their research and avoid repeating it. This could also be avoided by sampling the algorithms used for driving the cars with a larger and more exhaustive data-set. It is only a matter of time before we can find a solution to such an issue. But until then, one might question the safety issues involved with the automation technology used by various firms in its current state.