idea Owners
Motivation for the concept is the observation that bicycles (both self-powered and pedelecs) are the future of urban transport alongside (self-driving) electric cars, busses and rail services. In central London, for examples, bicycles already account nearly 1/4 of rush hour traffic-- in cities such as Copenhagen 41% of the population bicycle to work or school. Improving infrastructure to make cycling safer, more accessible and enjoyable are increasingly urban planning goals. To increase the attractiveness of cycling I propose the development of “smart” green waves for bicyclists. Green waves not only make cycling more efficient and attractive but also empirically reduce the likelyhood that cyclists will endanger themselves by running red lights. The last few years has seen a incredible wave of advances in machine learning and my idea is to apply these to address the needs and problems of urban transport. Instead of fixed "green wave" timings or priorities the idea here sets to increase the flow of bicycle traffic while also minimizing the impact on other traffic actors-- and in many use cases also resulting in improvements in general traffic times. Using low power efficient SoCs (Systems on a Chip) the "smarts" can be integrated in current and solar powered traffic lights. Using AI and computer vision allows for the development of systems that are not only more intelligent but also much cheaper to deploy than inductive loop, magnetometer or radar based sensors which need to be buried in the pavement.
The concept applies deep learning to traffic control with a focus of “smart” green waves for bicyclists. The solution deploys among other “technologies” the use of computer vision and pattern recognition to identify bicycles—no need for inductive loop sensors (which have a number of problems including non-detection of composites and false detection, for example, through neighboring large objects)—and pedestrians as well as a few novel applications of deep learning to try to reduce wait-time, increase the flow of bicycles while also minimizing the impact on other traffic participants. The system proposed is also compatible with other initiatives such as V2I (vehicle to infrastructure where the traffic lights provide information to the driver on impending signal changes). It can also be augmented by RFID and cellphone/ANT/bluetooth-- here registered phones can provide additional information into the system as well as provide information back to the cyclist about timings (simple traffic light countdowns can't be used due to the dynamic adaption of signals). Instead of fixing green-waves at some fixed speed (such as 20 km/h in Copenhagen, 13 mp/h in San Francisco etc. ) or simplistic priority driven traffic signaling, the system “learns” and is highly dynamic. It views traffic as part of a multi-agent strategic game exploiting the information about the movements of bicycles and motor vehicles. Traffic lights coordinate with one another to try to reach an optimal state according to policy. It learns and discovers "on its own" features of "efficient trafffic light control" rather than follow a crafted model.. Since the system has data on the rough number of bicycles, autos etc. as well as timings one can evaluate and adjust policies as new insights and urban transport goals develop.
The concept has been designed to meet (or exceed) German data privacy regulations-- Bundesdatenschutzgesetz (BDSG)-- which are amongst the strictist privacy regulations in the world. It learns features and nothing that can in the remotest sense be considered personally identifiable information.
In my future traffic lights will contain cameras and AI hardware-- GPU SoCs (System on a Chip) and/or ASICs. Currently developing on a NVIDIA Jetson TX1-- a development board for the Tegra X1 SoC (octa core processor with 4x ARM Cortex A57 cores, 4x ARM Cortex A53 cores, and a 256-core Maxwell GPU). Cost of the hardware is a fraction of the current costs, for example, alone of installing inductive sensors into bicycle paths . They are also getting more powerful—NVIDIA has already announced a new SoC for 2018 that will include the power of the current Drive-PX-2 board on an SoC. With the TX1 I’m getting adequate performance: well under ½ second to completely process a frame—real-time better than 2 fps—and identify objects and their spatial locations while comsuming only around 10 watts of power! With new inferencing hardware and software the kind of networks and their performance is getting better all the time—even on the TX1 I’m looking at getting a 2x speed-up within the next few months.
Since the detection of bicycles is vision based it can also detect abuse of bicycle lanes by other actors such as motorcars, delivery trucks as well as collect data on pedestrians crossing into lanes—as well as, of course, bicycles going the wrong way on one-way lanes.
Comments
Hi,
do you expect there might be opposition against such concept, because of privacy protection? There would have to be a camera with image recognition on every streetlight, just to create green wave for bicycles, that can easily stop and for example push a simple button. Do you think that the investment and maintenance cost would be covered by the benefits?
Maybe you are are just too zoomed in. What about connecting your image processing device to already existing cameras in frequent squares and crossings and control traffic lights and public transport in order to prevent overcrowding and traffic jams? Similar concept, but just applied on a larger scale.
Just some thoughts that crossed my mind, I don't mean to criticize :)
Jan
Privicy and data protection are highly regulated in Germany and the concept is designed to be wholly compatible. Cameras and AI hardware? Cost are, as I'm mentioned, much lower than other sensors currently in use. They don't need terribly high resolution or quality. The cameras are the lowest cost item in the chain- camera modules can cost as low as 20 EUROs. Using existing CCTV cameras I strongly suggest would be significantly more expensive and beyond cost pose a number of problems.
Prevent overcrowding and jams? Sure I see a number of applications such as controlling lights to empty gridlock. Would be relatively easy. V2I and smart traffic lights are, I think, on the table. The best solution, however, to attack gridlock I suspect is legistatiion and not just technology. If anyone caught in a gridlock can be fined a painful amount people would, as in California where such lesgislation was put into place, be less eager to intentionally block traffic..
Please note that the picture demonstarting object recognition above was made using a widely published press photograph of a bicycle path in Amsterdam-- published by the Amsterdam city hall as part of their dosier on cycling in Amsterdam-- rather than from real data as to not infringe on anyones privacy through publication. The recognition and identification is real.
Image source: https://www.iamsterdam.com/en/media-centre/city-hall/dossier-cycling
Can you please elaborate further the sustainability aspect of this solution? What benefits does it deliver in context of sustainability and do these benefits outweight the environmental costs of manufacturing and maintenance of the system?
The comments are closed.