starwit/overhead-traffic-anomalies

Overhead Traffic Anomalies (OTA)

• Developed by: H. Lichtenberg, Starwit Technologies GmbH, 2025
• License: Creative Commons Attribution Non Commercial Share Alike 4.0
• Master Thesis: H. Lichtenberg, Anomaly Detection in Traffic Applications: A Probabilistic Forecasting Approach Based on Object Tracking, 2025

This dataset contains traffic anomalies from static overhead cameras (intersections/roundabouts). It includes 32 anomaly categories with a relevance mapping for severity-aware evaluation. For each camera, we provide video frames with YOLOv8 detections, including 24 h of training footage and 48 h of test footage with 1027 anomaly annotations.

Please see master thesis for details.

This project was made possible by the city of Carmel, Indiana. Thanks to the mayor and the citizens of Carmel for supporting scientific research. Special thanks go the city's IT department for technical support as well as helping to understand, how traffic in Carmel works :) To find out more about Carmel look here: https://www.carmel.in.gov/

What’s inside

• Three camera scenes (each with traindata and testdata).
• Frames (320×180) as WebDataset .tar shards (frames-xxxxx.tar) plus a global index.csv.
• Object detections (YOLOv8) with tracking IDs and geo-coordinates in object_detections.json.
• Anomaly annotations in anomaly-labels.csv (test split only).
• Label dictionary in event_labels.txt and relevance mapping in relevance_mapping.json.

Anomaly labels

• label ∈ {-1, 0, 1, …, 32}
  • -1 = marks detection/tracking artifacts so they don’t bias evaluation (e.g., spurious box, ID switch)
  • 1…32 = anomaly categories (see event_labels.txt)
  • unlabeled trajectories are treated as normal (label 0)
• Relevance degrees (0–4) for severity-aware evaluation in relevance_mapping.json. The category→relevance mapping is subjective and may be adapted to match different application priorities.
  • 0: FP/uninteresting
  • 1: rather uninteresting anomaly
  • 2: relevant anomaly
  • 3: high relevance
  • 4: critical relevance (dangerous behavior)

Example anomaly categories

• Wrong-way driving (IDs: 22, 23) — vehicle travels against permitted direction
• Fast driving (ID: 11) — reckless speeding relative to scene context
• Traffic tie-up (ID: 16) — blockage or standstill due to congestion/obstruction
• Cutting off another vehicle (IDs: 18, 19) — failing to yield / forcing an agent to brake
• Broken-down vehicle (ID: 25) — stationary/disabled vehicle on a public road.
• Getting off the road (IDs: 28, 29) — getting off the roadway / parking on sidewalk

Wrong-way driving (23)

Cutting off (almost collision) (19)

Parking on sidewalk (29)

Broken-down vehicle moved by people (25)

File formats

• Frames: WebDataset shards frames-*.tar with JPEGs named frame_<timestamp_ms>_<running_index>.jpg.

• index.csv: maps each frame to its timestamp and shard, enabling alignment.

• object_detections.json (per scene/split): array of per-timestamp records with timestamp (UNIX ms), frame_index, frame_key (JPEG filename), shard (tar file), and a detections list. Each detection has class_id (YOLOv8), object_id (stable track ID), longitude/latitude, boundingbox normalized to [0,1], and confidence. The object_id persists across frames, enabling trajectory-level analyses.

• anomaly-labels.csv (test only): CSV with columns object_id,start_timestamp,end_timestamp,label. Contains labels for true anomalies (1–32) and input errors (−1).

Intended use

• Tasks: anomaly detection; severity-aware ranking; robustness to detection/tracking noise.
• Metrics: NDCG (severity-aware ranking); AU-PR, AU-ROC.

Quick start

Some download possibilities.

# Full dataset (all scenes + train/test + mappings)
from huggingface_hub import snapshot_download

snapshot_download(
    "HannaLicht/overhead-traffic-anomalies",
    repo_type="dataset",
    local_dir="OTA"
)

# Only test data (with anomaly labels)
snapshot_download(
    "HannaLicht/overhead-traffic-anomalies",
    repo_type="dataset",
    allow_patterns=[
        "MononElmStreetNB/testdata/**", "RangelineS116thSt/testdata/**",
        "RangelineSMedicalDr/testdata/**", "event_labels.txt", "relevance_mapping.json",
    ],
    local_dir="OTA-test-only"
)

You can align frames, detections, and labels using index.csv timestamps and object_ids.

import pandas as pd

# Example: load test annotations for a scene
root = "MononElmStreetNB/testdata"
labels = pd.read_csv(f"{root}/anomaly-labels.csv")  # object_id, start_timestamp, end_timestamp, label
index_df = pd.read_csv(f"{root}/index.csv")         # frame_key, timestamp, shard

# Get all frames within an anomaly interval (timestamps are UNIX ms)
def frames_for_interval(ts_start, ts_end, index_df):
    return index_df[(index_df["timestamp_utc_ms"] >= ts_start) & (index_df["timestamp_utc_ms"] <= ts_end)]

rows = frames_for_interval(labels.iloc[0].start_timestamp, labels.iloc[0].end_timestamp, index_df)
print(rows.head())

If you prefer streaming frames from .tar shards, consider the webdataset library.

License & Attribution

Data © 2025 Starwit Technologies GmbH - Licensed under CC BY NC SA 4.0.

Please cite:

H. Lichtenberg, Overhead Traffic Anomalies (OTA), Starwit Technologies GmbH, 2025

Acknowledgments:
Created by Starwit Technologies GmbH.
Detections use YOLOv8; geo-mapping via the Starwit Awareness Engine. Data was provided by the city of Carmel, Indiana