Traffic sign recognition on Android with reduced-complexity CNNs

The goal was a traffic sign recogniser small enough and fast enough to run on a phone, with no server and no network connection.

Three classification architectures were trained and compared on two datasets, GTSRB and BTSD, optimising for accuracy, response time and memory at once rather than accuracy alone.

• NL-CNN: 99.64% accuracy on GTSRB, 99.3% on BTSD, 0.22 million parameters, 0.73 ms response time. The best result of the three, and by far the most compact: published models with comparable accuracy are dozens of times larger.
• MicronNet: 97.62% on GTSRB, 98.28% on BTSD, 0.51 million parameters, 0.81 ms.
• MobileNetV3-Small: 95.76% on GTSRB, 97.11% on BTSD, about 1 million parameters, 1.94 ms.

Detection was handled separately, with the compact YOLO variants available at the time: YOLOv4-Tiny, YOLOv5s and YOLOv5n. GTSDB alone is skewed towards distant signs, so it was combined with a second dataset covering mid and near range. YOLOv5, with 1.76 million parameters, reached mAP@.5 of 98.4%, mAP@.95 of 68.4%, precision 97.4% and recall 95.7%, at 9 ms per 640×640 frame.

The models were exported to TensorFlow Lite and integrated into an Android application with two entry points: live recognition from the camera, and recognition on an image picked from the device.

Everything runs on the device. The upside is that no connection is needed and no data leaves the phone; the constraint is that the models have to stay small enough for the hardware, which is what drove the whole comparison above.