UTA for CS 381 Analysis of Algorithms

• Optimized real-time object detection on RTSP Raspberry Pi camera feed to run at >30 FPS

• Doubled YOLOv8 pose detection speed by refactoring C++ detection result postprocessing

• Wrote platform-agnostic benchmarking scripts for inference on Coral Edge TPU, Intel NPU, and CPU

Tutored students in Data Structures & Algorithms, Python, Calculus I & II, and Linear Algebra.

• Improved quantized YOLOv8 mAP by 7% through adjusting the model architecture to minimize concat layers

• Integrated a C++ postprocessor for YOLO detection into official DeGirum PySDK, running at <1ms per frame to parallelize with on-device inference with edge AI accelerators

• Trained models for 5 detection tasks, and exported to 160 total model variants

• - Datasets: COCO, License Plate, Face, Hand, and Car detection

• - Formats: yolov5nu/v5su/v8n/v8s, with SiLu/ReLu6 activations, to ONNX/TFLite/N2X

• Hosted a live seminar on YOLO quantization