jwt625/bay-bridge-traffic-cam

Bay Bridge Traffic Camera

Exact and derived time-series outputs from an experimental window-camera detector observing the San Francisco-Oakland Bay Bridge from August 2025 through July 2026.

These are algorithmic detections, not official or ground-truth traffic counts. Precision, recall, false-positive rate, and false-negative rate are unknown. Pixel-speed fields are detector features, not mph or km/h.

Contents

• raw/<metric>/<YYYY-MM>/part-00000.parquet: 108,644,312 exact Prometheus application samples with millisecond timestamps, raw values, and original labels.
• derived/browser-v3/: compact five-minute, half-hour, hourly, daily, profile, coverage, lighting-regime, and analysis tables used by the static site.
• validation/external-v1/: comparisons with MTC and Caltrans high-level reference statistics, including provenance.
• exact-export-provenance.json: export boundaries, image identity, schema, and row counts.
• manifest.csv: checksums for the complete published package.

Native Prometheus TSDB blocks and the Grafana database are deliberately not published because they can contain operational metadata or credentials. No continuous video archive was collected.

Exact raw schema

Each Parquet shard uses the same schema:

• timestamp_utc: timezone-aware millisecond timestamp
• metric: Prometheus metric name
• value: unmodified floating-point sample value
• labels_json: every original Prometheus label, sorted as JSON
• typed convenience label columns: direction, window, component, app, instance, job, exported_job, exported_instance, and exported_exported_instance
• source_series: direct or imported

Historical import repairs created duplicate-looking label series. The exact layer preserves both. For most traffic analysis, select source_series == "direct" or use the already reconciled derived tables.

Row inventory

Metric	Exact samples
motion_detector_fps	5,917,120
system_status	17,755,458
tracked_objects_active	5,917,120
traffic_flow_rate_per_minute	11,836,130
traffic_speed_average_pixels_per_second	32,661,438
traffic_speed_current_pixels_per_second	10,883,256
traffic_vehicles_created	11,836,796
traffic_vehicles_total	11,836,994

Loading examples

import pyarrow.dataset as ds

flow = ds.dataset(
    "raw/traffic_flow_rate_per_minute",
    format="parquet",
    partitioning=None,
)
table = flow.to_table(
    filter=ds.field("source_series") == "direct",
    columns=["timestamp_utc", "direction", "value"],
)

Or after publication:

from datasets import load_dataset

flow = load_dataset(
    "jwt625/bay-bridge-traffic-cam",
    data_files={
        "train": "raw/traffic_flow_rate_per_minute/**/*.parquet"
    },
)

Interpretation and known discontinuities

The two camera directions have materially different view and occlusion responses. A comparison with 2024 Caltrans AADT suggests indicative presentation multipliers of 2.447544× for Oakland-bound (left) and 1.356866× for SF-bound (right). These are reversible reference-scaling factors, not calibration constants. Published raw and derived files remain unscaled.

Bay Lights LED commissioning began approximately 2026-02-19, followed by the official relighting on 2026-03-20. Nighttime detector spike excess and within-night variability increased by roughly 3×, consistent with animated lights being detected as motion. Use the pre_lights, commissioning, and illuminated fields to separate optical regimes; do not interpret the post-lighting nighttime surge as traffic growth.

Coverage gaps, one observed direct-counter reset, DST handling, external validation, and detailed methodology are documented in the project repository: https://github.com/jwt625/bay-bridge-traffic-cam

License and citation

Data and derived tables are released under CC BY 4.0. Code is separately released under MIT. Suggested attribution:

Jiang, Wentao. Bay Bridge Traffic Camera Dataset (2025-2026).