Project highlights
-
Custom Model Training: The YoloV8 model was fine-tuned with a dataset of over 10,000 Colombian license plates, considering regional variations and deterioration. This specialized training resulted in a 30% increase in recognition rates compared to generic models.
-
Real-time Processing: The software processes video feeds at 20-30 frames per second on standard hardware, with intelligent frame sampling to balance accuracy and performance. A buffering system ensures that no plates are missed during processing spikes.
-
Character Recognition Enhancement: Advanced post-processing algorithms correct common OCR errors specific to license plate fonts, improving character recognition accuracy from 85% to 98%. The system can distinguish between similar characters (e.g., 8/B, 0/O) with high confidence.