Cylab Security & Privacy Institute

Jun, 2025 - Present  •  1 yr 3 mos  -  Part time

• Designed a real-time multimodal perception pipeline for autonomous driving to detect flashing
emergency vehicles and traffic lights under strict latency constraints.
• Gathered a 63K-image dataset and fine-tuned YOLOv12 and ViT-B/16, achieving up to 97% mAP for
emergency vehicle detection and 99% traffic light and 95% emergency flash classification accuracy.
• Optimized end-to-end inference via TensorRT quantization, reducing total pipeline inference latency
from ~88 ms to ~18 ms (4.9x speedup) while preserving model performance.