David

Algorithm Engineer

Aug 2023 - Aug 2024 • 1 yr

Background: Real-time workplace safety monitoring for drowsiness, climbing, overstepping, smoke, and flame detection in workshops/offices. Technical Approach: • Drowsiness/Climbing/Overstepping: YOLOv8-pose, refactored two-stage (detection + pose-crop classification) into single-stage pose classification with Pose branch as auxiliary loss for Backbone training. Effectively distinguished target actions from confusable ones (tying shoelaces/bending over), with temporal filtering (alert after 10 consecutive frames). • Flame/Smoke: YOLOv8 object detection with 6,183 flame and 54,000+ smoke samples, custom-trained for Baoji site to address small/thin smoke misses and night insect-trail false positives. Key Achievements: • Drowsiness: mAP@50 98.8%, precision 98.3%, recall 98.4% – passed on-site customer acceptance. • Flame: Delivered YOLOv8m model resolving false positives from red paint, warning lights, and similar flame-like objects.

Algorithm Engineer

Feb 2022 - Jul 2023 • 1 yr 5 mos

Background: Financial-grade ID/driver's license OCR with privacy constraints. Technical Challenges: Label scarcity, skewed/multi-line text, rare/traditional/similar Chinese chars. Technical Approach: • Data Synthesis: Custom generator based on Chinese Coded Char Set (8,300+ chars), producing 10M+ synthetic samples. • Model Selection: Compared SVTR vs CRNN, adopted single-line detection + classification head to resolve multi-line ambiguity. • Post-Processing: CTC decoding with dictionary constraints to filter invalid combos, boosting full-field accuracy. Key Achievements: • ID Card V3: Detection P/R 100%; Recognition P/R 87.12%; Char accuracy 98.19%. • Final SDK: Optimized for similar chars and special symbols (e.g., "·", "‑") in minority IDs, delivered C++ SDK with sustained full-field accuracy improvements.