Let's shed some light on how Telraam's night counts differ from its daytime operations, focusing on the unique challenges and adaptations necessary for nocturnal data gathering. Read more in this FAQ article about how we measure at night, the accuracy, the single class categorisations, and the speed measurements.
How do we count during the night?
Telraam uses the night count algorithm when it is dark outside. This differs from the day counting algorithm. While when it is bright outside the AI chip can identify and classify various road users, and track them as they move along the street, when there is no sufficient ambient light the only thing the AI can do is tracking moving light sources. On dark streets the only thing the camera of the Telraam S2 device can see is the light streaks of vehicles passing by, and the night model of the AI chip counts (based on) these.
Accuracy
Achieving the same level of accuracy as daytime operations is a tall order under the cover of darkness, mainly because the (front and rear) light visibility of different vehicles vary depending on their make and model, the placement of the Telraam S2 device, and other variables. However, after testing and diving into the statistics of the manual counts of our beta-testers, we have at this moment an accuracy rate of around 70% for the night counts. We will update our general accuracy documentation on GitHub regarding night counts when we have enough validation data to do so.
Single class
During the night all traffic is aggregated into a single category. Unlike during the day, where Telraam can distinguish between up to 10 modes of transport, night counts may include any type of vehicle that has strong lights. In practice this means that night counts are an aggregation of motorised vehicles. This simplification stems from the way how Telraam S2 counts at night as explained earlier.
No speed measurements
With the current night count algorithm it is not possible to determine speeds of the motorised vehicles. During the daytime speeds are estimated using the apparent length of vehicles classified as cars, but when the vehicles themselves are not visible then we have no means to identify cars safely as cars, thus we have no scale reference for this kind of estimation. We found that trying to use both the front and rear light of vehicles is way too inaccurate for this kind of work, as first of all rear lights are often not visible at all, or not visible at the same time as the front lights, and in the dark we cannot safely tell if these lights are connected to a motorbike, a car, a van, or a truck, making class-based speed-estimations currently impossible.
Roll out
Development of the night counting algorithm started around January 2024, and we carried out a first test using Telraam S2 test devices in our office around mid February 2024. After a positive evaluation, we went to the next stage and deployed the night count algorithm to 24 beta-testers. These users were also asked to do a manual validation count during day- and night time to evaluate the accuracy of the algorithm, and to see if any unexpected problems would appear.
Currently, we find ourselves in a meticulous phase of refinement, mainly addressing a specific issue that caught our attention during these evaluations. Our commitment to excellence means that we are taking the time needed to ensure everything is just right. Once this hurdle is cleared, we are excited to roll out the new firmware to all Telraam S2 devices, bringing enhanced capabilities to our community and paving the way for round-the-clock traffic insights.
More questions on the night counts? You can contact us via support@telraam.net
