01 — 2025 — present
Edge AI Vision Deployment
Detect-and-Avoid perception on resource-constrained onboard computers.
- Role
- Undergraduate Research Assistant — ARC Lab
- Stack
- DarknetYOLOTensorFlowKV260FPGA-Based SoCPython
Problem
Deep learning detectors trained on workstations do not survive the move to onboard compute. The system has to detect aircraft in real time while accuracy, latency, power, and memory all compete for the same budget.
System
A standardized deployment path: training and export, post-training quantization, onboard inference runtime, and a benchmark harness that reports accuracy and latency from the same run.
- Training & export — model checkpoints exported to a deployable graph
- Quantization — precision reduced with paired accuracy tracking
- Onboard runtime — real-time inference on the embedded target
- Benchmark harness — mAP, precision, recall, and FPS measured together
My contribution
- 01Built the deployment pipeline end to end, from training output to onboard inference.
- 02Performed model quantization and latency/accuracy benchmarking on the embedded target.
- 03Collaborated with CV researchers to integrate newly trained models into the pipeline.
- 04Standardized the deployment workflow so later iterations reuse it without rework.
Engineering decisions
Stable embedded deployment
Problem
Deploying models onto resource-constrained onboard computers while holding real-time inference.
Decision
Built one standardized pipeline covering training, quantization, deployment, and benchmarking instead of per-model scripts.
Tradeoff
Higher upfront engineering effort.
Outcome
Significantly improved reproducibility across model iterations.
RGB vs grayscale
Problem
Grayscale input cut computational cost but unexpectedly caused a large accuracy drop.
Decision
Designed paired RGB / grayscale experiments to isolate image modality as the single variable.
Tradeoff
RGB 65–70% mAP vs grayscale 30–40% mAP.
Outcome
Kept RGB in the deployed configuration; root cause still under investigation.
Results
Measured benchmarks
Build status
- Reusable deployment pipeline
- Quantized onboard build
- Benchmark harness
Build status — not a measured benchmark.
Lessons learned
- Embedded AI is a simultaneous balance of accuracy, latency, and hardware limits — the right tradeoff depends on the use case, not on a general rule.
- Standardized pipelines remove most of the repetitive integration cost between model iterations.
- Deployment is a system problem spanning preprocessing, inference, post-processing, and hardware — not just model optimization.
Next project