This article is part of a series of articles that showcase the real world applications of IoT. Keep watching this space for more.
Reading the newspaper, I came across an article which said Singapore has begun its first trial for driverless taxis. We have been hearing for a while that driverless cars are coming and there have been numerous hurdles from regulatory and safety angles. However, it seems that they have been overcome to some extent to allow this trial to commence. This blog examines in brief how cars can drive themselves and how analytics is a vital enabler in this seismic shift in the automotive industry.
Many of us may recall how Deep Blue, in May 1997, an IBM computer famously defeated the then world chess champion Garry Kasparov over six games. You may wonder what this has in common with driverless cars. Both employ a form of machine learning called deep learning. Though my understanding of this technique is limited, I will attempt to unpack it a little to understand some of the basics of what drives a driverless car.
All vehicles, irrespective of size, can be fitted with sensors/cameras etc., i.e. devices that relay data back to a system. This data can be processed before it hits the database, which provides real-time analysis on what is commonly referred to as event stream processing. The analysis can be a very simple rule-based processing or can take a more sophisticated form. With driverless cars, the analysis is a technique called deep learning. Some refer to it as a rebranding of neural networks; but essentially, it is a technique that involves several layers of largely unsupervised or semi-supervised learning. This methodology has been fine-tuned over the years in areas such as Deep Blue, speech recognition, and text mining and is now being adapted for more advanced usage and the case being discussed in this blog, autonomous cars.
The Swedish automotive superpower Volvo will launch their pilot program – Pilot Assist II – in association with Uber in 2017 and will render their cars capable of steering and braking up to a speed of 80 miles per hour. While it does not completely do away with the need for a driver, these are significant strides in that direction.
The objective of deep learning is to allow an algorithm to detect or learn various parts of a car such as wheels or other components of a car, objects/persons travelling/present around the driverless car and as it is exposed to more inputs, connections begin to develop to the eventual outputs and the actions to be triggered in the car in response.
In a pedestrian detection algorithm that is currently being developed by many, including Google, the first step is to weed those parts of an image which are obviously not people, using simpler techniques before employing the complex deep learning algorithm that this application calls for. As of six months ago, this technology was not sophisticated enough to allow for real-time detection that a completely autonomous car would call for, as the more sophisticated algorithms require more time for processing which is then a safety hazard. As of now, since safety is the number one priority, this algorithm in not being implemented.
While there are some considerable challenges that need to be surmounted before the fully autonomous vehicles are launched, many of the big investors in this technology look to be on track to meeting their own 2020 deadline to launch fully autonomous cars. To be successful, the key is the ability of technology to deal with the exceptions in a more predictable fashion. The launch of trials of these scaled down versions or those with limited features will allow developers to collate some valuable data to further fine-tune the algorithms.
Susan Mani, Analytics Expert
Susan is a seasoned analytics professionals with over 10 years of experience working on analyses for Fortune 500 clients such as Bank of America, Proctor & Gamble, and Unilever among others. She is on the Alumni Advisory Board of the University of Cambridge, and is an expert in the practical aspects of applying analytics in different contexts to generate value.