Providing Drivers a Safety Net with Computer Vision
This year, the National Safety Council projects that more than 30,000 people will be killed and 400,000 injured as a result of distracted driving
While automotive safety features and driver assistance technologies have successfully lowered fatality rates, the ever-growing distractions of modern life such as smartphones and infotainment systems have proven stubbornly difficult to overcome.
Automated driving technologies have the potential to save lives and there is a great desire to get these systems on the road as soon as possible. State and municipal governments have been especially accommodating to pilot programs, and millions of miles have already been driven autonomously on America’s roads and highways. Most industry experts and analysts agree that autonomy is rapidly evolving, but implementation and development of the technology have faced numerous roadblocks.
Like many nascent technologies, the autonomous drive has entered the market sporadically via beta versions and minimally viable products that will achieve their full capabilities only after years of real-world testing and improvements. With potential lives at risk, testing requirements are more rigorous in comparison to other emerging technologies.
A primary reason for delays is planning and design for different environments. Engineers and product managers must develop for highly congested cities as well as rural areas, with various environmental conditions and zoning requirements. Suburban areas with well-maintained infrastructure and fair weather lend themselves well to autonomous drive, but rural and crowded urban areas may need more planning in the long term. It is for this reason that we see Google and Uber running self-driving pilot programs in places like Phoenix, Arizona.
Limited deployment in geographically restricted areas with minimal weather variation seems like the standard environment for initial pilots of autonomous drive. However, this strategy is only a stepping stone on the way to a national rollout.
The biggest roadblock to full autonomy is that Artificial Intelligence (AI) systems are data hungry and their performance is correlated with access to millions of carefully labelled examples to mimic real-world features. A dizzying array of environmental conditions and edge cases must be engineered in order to be robust enough to reflect the real world, from geographic features and landmarks to living beings. In addition, data must reflect all potential sensor modalities, including visible light, ultraviolet, infrared, radar and LiDAR.
Semantic segmentation or analysis at the pixel level is notoriously tedious and expensive for autonomous drive. Accurate labelling can take up to 1 hour per image and each segmented image can cost up to $10, which quickly becomes an astronomical cost given the millions of images required.
Unsurprisingly, major companies have led the way in autonomous vehicle tech, with Google’s Waymo riding 5 billion miles in the simulation last year. The company has created detailed simulations of cities like Austin, Texas, Mountain View, California, Phoenix, Arizona and others. Photorealistic simulations were designed to contain tens of thousands of virtual automobiles and pedestrians and allow for precise control of environmental, road conditions and edge scenarios. The simulations also allowed for the testing of corner cases such as car crashes, pedestrian accidents, natural disasters, catastrophes and ethical dilemmas that would neither be possible or ethical to run in real life.
While Google simulates autonomous drive in a few lucky areas of the country, there are few firms with the resources to leverage the technology to drive change in other areas. Autonomous driving systems can be applied to applications such as industrial inspection, autonomous mining, precision agriculture, marine navigation and operations, forestry, search and rescue, emergency services, wildlife conservation, law enforcement, package delivery, robotic food service and much more. The possibilities are endless and the scope of these applications is truly too large for any of the current crops of autonomous driving companies to tackle alone.
With the rise of computer-generated or Synthetic Data, companies at all scales can easily develop the necessary data assets to power leading-edge AI and Deep Learning applications at a fraction of the time and cost of externally acquiring and hand labelling training data. Recently noted in TechCrunch and Wired as a disruptive force in AI, Synthetic Data significantly reduces the difficulties associated with the preparation, labelling and collection of data for AI applications.
Synthetic Data is fast becoming an essential component of autonomous driving and computer vision AI systems. By bringing together techniques from the movie and gaming industries (simulation, CGI) together with emerging generative neural networks (GANs, VAE’s), we are now able to engineer perfectly-labelled, realistic datasets and simulated environments at scale. There is virtually no incremental cost of additional generated images and since the Synthetic Data is created all the attributes are known to pixel-perfect precision. Key labels such as depth, 3D position and partially obstructed objects are all provided by design. Application of this technology could allow important safety features to be brought to market quickly and cost-effectively, from crash prevention software to predictive maintenance, onboard diagnostics, and location insights.
Synthetic Data is a cost-effective solution that cuts down on the time and effort needed to acquire, clean and organize driver data. It is infinitely scalable and can be diversified to reflect the behavioural and demographic attributes of real-world populations – reducing bias and eliminating the need to navigate cumbersome regulatory hurdles. Even given the increased implementation of sensor applications, hardware, and cloud connectivity in electric and autonomous vehicles, usable driver data may not be readily accessible to developers – due to heightened privacy concerns. Unlike organic data, Synthetic Data can be customized for the specific road safety learning and testing applications and interfaces of the modern connected vehicle.
All credits to the source below by Yashar Behzadi