Edge AI / Senior capstone · In progress
Task-based PPE detection
Comparing mask and footwear classifiers as a baseline for task-aware workshop PPE monitoring.
01 / CONTEXT
The challenge.
Detecting a glove or a pair of glasses in a frame does not establish whether a technician is wearing the right equipment for a task. Small items, occlusion, lighting, and camera angle make automotive workshop evaluation difficult.
02 / PROCESS
The approach.
- Define task-to-PPE rules with workshop staff, beginning with one repair bay and one camera.
- Separate equipment presence, association with a worker, task relevance, and uncertain or occluded views.
- Develop a public-data classification baseline comparing ResNet-50, EfficientNet-B0, and YOLOv8n-cls for mask and footwear regions.
- Plan independent workshop evaluation with class-level precision, recall, error counts, and measured camera-to-result delay.
03 / REFLECTION
The outcome.
Completed six public-data classification runs on a Tesla T4, comparing ResNet-50, EfficientNet-B0, and YOLOv8n-cls. Validation selected ResNet-50 for masks, with 96.55% test accuracy across 87 crops, and EfficientNet-B0 for footwear, with 100% across 120 crops.
Project scope
Public-data region classification baseline, using 3 inferred test groups for masks and 4 for footwear. Temporal blocks can share recordings or people across splits. These results describe crop classification; workshop evaluation and end-to-end edge deployment are the next development stage.
04 / RESOURCES