Project Detail

Road Safety Impact Evaluation (Addis Ababa)

Built a full video-to-metrics pipeline for road safety impact evaluation, from object detection/tracking in intersection footage to conflict/traffic analytics and reporting.

Problem

Manual video review could not scale across many intersections and time periods, making it difficult to produce consistent conflict and traffic indicators for baseline/follow-up evaluation.

Approach

I implemented and operationalized a two-stage workflow. Stage 1 (video processing): ran YOLOv7 + DeepSORT on MP4 intersection footage to detect and track road users, and produced annotated outputs plus metadata (JSON/parquet), with batch execution and storage to S3. Stage 2 (data processing): transformed per-location parquet outputs into analysis-ready datasets using movement polygons, direction splits, geolocation/homography metadata, and baseline/follow-up logic; generated conflict and traffic aggregates, trajectory plots, short reports, and one-pager/dashboard-ready artifacts. I also supported subsampling/bootstrapping analyses to evaluate metric stability when using partial data.

Results

Delivered reproducible, scalable safety analytics across locations, reduced manual processing load, and enabled longitudinal traffic/conflict comparisons for impact evaluation reporting. The related World Bank paper is now published. The codebase is also being packaged for reuse in other World Bank projects. Read the paper.