SafeSight: A Scaled Prototype for Semi-Supervised Activity Recognition, PPE Vision Detection, and Retrieval-Grounded Safety Reporting

This report presents a small-scale prototype exploring three components relevant to sensor- and vision-based occupational safety monitoring: (1) semi-supervised activity classification under simulated label scarcity, (2) image-based personal protective equipment (PPE) compliance detection, and (3) a minimal retrieval-augmented generation (RAG) pipeline for producing guideline-grounded safety notes from classification outputs. Using the public UCI HAR dataset, a logistic regression baseline trained on 10% of available labels achieved 92.8% test accuracy; a confidence-thresholded pseudo-labeling scheme did not improve on this baseline at any tested label fraction (2%, 5%, 10%), yielding a small but consistent decrease in accuracy. For the vision component, a ResNet18 fine-tuned on cropped head/helmet regions from a public hard-hat detection dataset achieved 99.45% validation accuracy, balanced across both classes. A lightweight TF-IDF-based retrieval step was then used to ground classification outputs in short safety-guideline text, demonstrating a complete classification-to-report pipeline. This work is exploratory and scoped to public benchmark datasets over a short development period; it is intended as a technical report and portfolio artifact, not a peer-reviewed contribution.