Computer Vision (CV) models are algorithms trained to recognize patterns in images and videos and need to be frequently updated to maintain accuracy with evolving visual environments. License Plate Recognition (LPR) systems use cameras and CV to read license plate numbers on vehicles, aiding in identifying stolen vehicles, uninsured motorists, and more. Tapway, a leading AI/ML & IoT company, offers a vehicle profiling platform called VehicleTrack for LPR and vehicle classification.
To address challenges like unseen edge-cases and evolving data faced by LPR systems, Tapway implemented an Incremental Training Pipeline (ITP) for their CV models. The ITP ensures models are continuously updated and performance improved over time. The overall architecture of VehicleTrack involves edge devices, AWS services like SageMaker and Kinesis, and Amazon OpenSearch for data visualization.
The ITP architecture includes sections for collecting new images, data preparation for model training, ML training jobs, and model deployment. The system addresses scenarios like minority classes and unusual predictions by collecting new training data. Tapway’s ITP has enabled them to train more accurate models in a shorter time frame compared to traditional manual methods.
The blog post showcases examples of how Tapway’s ITP has improved model performance with successive iterations. Real-world edge-cases are presented with model predictions, demonstrating the robustness of the system. Deployed on multiple highways in Malaysia, Tapway’s LPR solution achieves high accuracy and continues to evolve through incremental training to tackle unique challenges.
Tapway has addressed new challenges faced after deploying the LPR solution, like occlusions and motion blur, by implementing techniques such as fine-tuning OCR parameters and implementing a voting algorithm for prediction selection. The ITP has enabled Tapway’s LPR and vehicle detection CV models to handle real-world conditions and adapt to changing data distributions, making them more robust and accurate over time.