There are all manner of hurdles to overcome before self-driving cars dominate our roadways, not the least of which are bicycles. As autonomous-technology developers continue to refine the software providing the capability for a vehicle’s array of sensors, cameras and LIDAR to work together in identifying objects surrounding them, one thing is clear: Autonomous systems are better at recognizing some objects than others.
IEEE Spectrum, the website for the largest professional organization for professionals in engineering and applied sciences, recently identified bicycles as a chronic issue for autonomous technology. Thanks to their agility, lack of mass and the unpredictability of their human operators, bicycles pose an even greater problem to autonomous vehicles (AVs) than to human drivers.
Oops, There It Is
Let’s face it: Bicycles have a very small footprint compared to most other vehicles sharing the road. They are easily lost against the backdrop of much larger objects, be they 18 wheelers, parked cars or even a hedgerow along the street. Moreover, most exercises in training autonomous systems to recognize objects have concentrated on motorized vehicles like cars and trucks, rather than bicycles. Given their size, lack of mass and the fact that they have been basically an afterthought in cataloging objects, it’s no wonder autonomous systems often misidentify them or overlook them entirely.
According to IEEE, in one test, the Deep3DBox algorithm — a new technology for detecting 3D objects — identified cars nearly 90 percent of time, but bicycles in less than 75 percent of encounters.
Another issue: Is that a motorcycle or a bicycle? Two-wheel vehicles can appear much alike to software interpreting camera and radar images. Bicycles and motorcycles have different capabilities and different rules of the road. Misidentifying a motorcycle as a bicycle, or vice versa, can have grave consequences.
Coming or Going?
Even when an autonomous system correctly identifies a bicycle, that’s only part of the issue. Nearly as important as properly identifying an object is determining its orientation. That is, which way is it facing? Obviously, determining an object’s orientation is made simpler if it’s in motion. But if it’s stopped at a traffic light or pausing along the side of the road, deciding which direction in which it might move can be problematic for autonomous systems.
Here again, IEEE reports that the Deep3DBox, which it identifies as one of the best, was able to correctly determine the direction of 88 percent of the cars, but just 59 percent of the bicycles.
Zigging or Zagging?
Autonomous systems will continue to improve in identifying and determining the orientation of bicycles, but yet one more issue looms in self-driving cars being able to coexist with bicycles on the road: those darn human riders.
One stumbling block to AV development casts a shadow over transitioning from human-driven to fully autonomous cars — and that’s AVs sharing the road with human-driven vehicles. Until every vehicle on the road is fully autonomous, those that are will have to predict what the human drivers will do. This is every bit as true with bicycles. AVs must be able to anticipate any move a human bike-rider might make. How far away from that might we be, when today’s best system can only determine the direction in which a bicycle is pointed 59 percent of the time?
What It Means to You: Indeed, AVs have a bicycle problem. It can and will be solved, but no one knows the timetable.