The good thing is that an AI model can learn the distinguishing features that really matter to control an intersection. FlowCube: capturing cyclists and pedestrians The FlowCube traffic sensor is used primarily for traffic at intersections. Traffic flows are captured using Vision AI. The great strength of the FlowCube is its flexibility. It can learn to recognize cyclists, pedestrians, and any other type of road user. By placing FlowCubes at multiple locations, the routes of individual road users can be determined – without identifying them or storing any data. In addition, we use the FlowCube to record near misses at busy intersections. The information we collect in this way can be a valuable contribution to monitoring the handling of traffic. In the Netherlands, the FlowCube has proven ideal for obtaining a better view of cyclists and pedestrians (motorized traffic can be ‘seen’ with induction loops). In the US, where intersections rarely or never have any detection infrastructure, cities are using the FlowCube to provide traffic control information about all modes of traffic – including public transport. AI and the smart TLC It can be more of a challenge to apply Vision AI in the Netherlands. There are many iTLCs in this country: smart traffic light controllers. iTLCs receive information about the traffic situation from traditional detection loops and vehicle GPS information. Using queuing models and optimization techniques, the traffic controller then determines the optimal sequence for green for each direction on the intersection. Vision AI applications such as the FlowCube can currently 22 only present their information to an iTLC if they have removed all contextual information concerning vehicles (distance to the stop line, length of the vehicle, lane), whereupon the traffic controller reconstructs the information that is relevant to its task. This means potentially useful information is being discarded, so that the power of Vision AI is not used optimally. Do more with less data So what can we expect AI and traffic management to do? What Large Language Models have proven is that the total corpus of information on the internet can be summarized in a relatively limited set of weights of a neural network. In technical language: the entropy of the data and the knowledge that is collected and stored is high, and AI models can make do with much less data. This is possibly true also for specific applications such as controlling traffic at intersections. AI is openminded, some would even say naïve. The good thing about this is that an AI model can learn the distinguishing features that really matter to control an intersection. People think that a model will have to be able to count individual vehicles in a queue, but maybe that isn’t necessary at all for good traffic control scenarios, for example if the model can see other useful patterns, such as the length of the queue, or the risk of a tailback. The next question is whether we humans are able to make sense of it and – crucially! – whether we have the confidence to allow a model like this to control the traffic. In any case, any model will have to be tested at length and thoroughly before we allow it anywhere near the real world. We don’t know yet what they will look like, but we do know one thing: we are going to see new AI applications in traffic management soon.
23 Online Touch Home