License Plate Recognition in Low Light
A quick experiment with the Ubicept Evaluation Kit
Introduction
One of the exciting things about having an evaluation kit is that our less-technical team members can quickly experiment with interesting use cases. Recently, one of our newer hires was inspired to push the limits of our solution after seeing our successful QR code demo. He wanted to know if our solution was able to perform automatic license plate recognition (ALPR) at night. We were thrilled with his curiosity and encouraged him to give it a try. After all, this was a great opportunity to test both the performance of our solution and the ease of use of our evaluation kit!
Conventional surveillance cameras are often unsuitable for ALPR because of their inability to provide blur-free images in low light. In addition, variable illumination from other light sources (e.g., headlights from oncoming vehicles) can make proper exposure very difficult. Here’s a still frame that our team member took from a 1080p pan-tilt-zoom camera with a low-light sensor:
As you can see, the results on moving vehicles are unusable for downstream ALPR tasks. Notice that the license plate on the parked vehicle is also unreadable despite the sensor itself having high enough resolution. This is due to the noise from insufficient light.
To be clear, this doesn’t represent the current state of the art in ALPR. Many specialized cameras are available today that use powerful infrared lights and high shutter speeds to capture clear images of retroreflective license plates. The question that we wanted to address was whether the Ubicept solution could accomplish this same feat quickly and without any supplemental illumination.
The Test
The camera was mounted on a tripod and equipped with a 75mm lens and 2x teleconverter to get a tight shot of the street approximately 50 meters away. Note that it was pointed out of an office window, so any illumination from the camera (infrared or visible) would have resulted in significant glare.
Using this incredibly simple setup, our team member was able to capture precise footage of all the license plates passing by. Here's some slow-motion footage of a single vehicle, with some information censored to protect the driver's privacy:
For the actual recognition task, we typically encourage customers to pursue API-level integration with our perception toolkit for maximum performance and efficiency. However, our team member (as someone who doesn’t write code) chose the alternative approach of sending the video directly from the evaluation kit to a third-party perception solution (Rekor Scout):
Summary
We want to emphasize that we’re not trying to criticize the low-light camera that we used in the comparison. As we mentioned, it simply wasn’t designed or advertised for ALPR, and there are specialized solutions that could match the performance we were able to achieve with the quick-and-dirty deployment of our evaluation kit.
Instead, we want to highlight the power and flexibility of the Ubicept solution to address diverse perception tasks with minimal effort. We used the same hardware (with different lenses) to address QR code detection, 3D reconstruction, high-speed person detection, extreme dynamic range for ADAS, and more.
We're confident that the Ubicept solution can help you solve your unique visual perception problems. If you want to learn more, please contact us, order an evaluation kit or set up a consultation today!
Updated March 18, 2023: We’ve made an extended version the video clip available below.