From 50% to 95% vehicle detection, right on the edge

DataCollect builds ARGOS, a camera-based traffic counting system designed for mobile, battery-powered deployment. Hundreds of units in the field across Germany. The problem: vehicle detection worked reliably only about 50% of the time. An earlier university partnership had not delivered the results they needed.

The issues ran deep. The Basler camera produced unusable images in poor lighting. The MobileNet V2 detector lacked the accuracy required for real-world conditions. And the NVIDIA TX2 hardware could not keep up with the computational demands of real-time processing.

100 DAYS redesigned the entire detection pipeline. We replaced the camera with a Sony IMX 327 for better low-light performance. We swapped MobileNet for a YOLO CNN, trained on a purpose-built dataset. For compute, we moved to the NVIDIA Jetson NX, significantly more powerful than the TX2 at the same power envelope. The full pipeline runs on PyTorch and NVIDIA DeepStream.

Our data scientists worked alongside DataCollect's internal team to systematically expand training data and iteratively improve model quality. Knowledge transfer was built into every step.

The result: 95% detection accuracy in field operations. Stable across hundreds of deployed units, in all weather and lighting conditions. DataCollect's team now trains new models and extends the system independently.