TTC logging + review dashboard

Turns the ttc_v2_label_node topics into a reviewable CSV + frame log, either from a finished rosbag2 recording or live during a demo, and a Streamlit app to step through the result frame by frame.

Layout

configs/
  classes.yaml       locked COCO classes (reference for the perception node)
  thresholds.yaml     TTC band thresholds used by label_logic.py
src/
  label_logic.py       pure fn: (valid, distance, ttc) -> (safety_label, potential_collision)
  csv_writer.py         appends CSV rows + saves debug images for a run
  extract_from_bag.py   offline: reads a finished rosbag2 recording
  live_logger_node.py   ROS2 node: subscribes live during a demo
dashboard/
  app.py                Streamlit viewer, reads a run folder only
data/runs/<run_id>/     one folder per run: log.csv + frames/

extract_from_bag.py and live_logger_node.py both call the same label_logic.compute_label() and csv_writer.RunLogger, so a bag replay and a live demo of the same input produce identical rows.

Assumptions to check against your actual node

• Message types: Bool / Float32 / Float32 / String (JSON) / Image for the five topics. If your node publishes custom message types, update MSG_TYPES / the subscription calls in both logger scripts.
• objects_json is a JSON list of objects with at least class and distance_m keys. primary_object_class() in each logger picks the nearest one as the row's obstruction_class. Adjust the key names if your schema differs.
• The five topics aren't assumed to publish in lockstep. ttc/debug_image is treated as the per-frame "tick"; a frame is only logged if the other four readings are all within --sync-tolerance seconds (default 0.1s) of the image. Frames without a fresh reading are skipped and counted, not logged with stale data.

Label bands (configs/thresholds.yaml)

condition	safety_label	potential_collision
valid = false	INVALID	(blank — unknown, not "N")
valid = true, TTC = inf	OBSTRUCT	N
TTC > 4s	CRUISING	N
2s < TTC ≤ 4s	WARNING	N
1.5s ≤ TTC ≤ 2s	POTENTIAL_COLLISION	Y
TTC < 1.5s	HARD_BRAKING	Y

All four numbers live in configs/thresholds.yaml, not in code.

Running it

Offline, from a finished bag:

pip install pyyaml opencv-python --break-system-packages   # inside your ROS2 env
python3 src/extract_from_bag.py /path/to/bag_dir \
    --thresholds configs/thresholds.yaml \
    --runs-root data/runs \
    --run-id 2026-07-03_demo1

Live, during a ROS2 session:

python3 src/live_logger_node.py --ros-args \
    -p thresholds_path:=configs/thresholds.yaml \
    -p runs_root:=data/runs

(Wire this into your package's setup.py console_scripts to run it as ros2 run <your_package> live_logger_node instead, once it lives inside your ROS2 workspace.)

Dashboard (no ROS2 needed):

pip install -r requirements.txt
cd dashboard
streamlit run app.py

It lists every folder under data/runs/, lets you filter by label and class, step frame-by-frame, and shows a TTC/distance chart for the whole run — useful for spotting near-miss events to pull out for edge-case review.

Tested so far

• label_logic.py: unit-tested all six band boundaries (exact 4.0s, 2.0s, 1.5s edges) plus the inf and invalid cases.
• csv_writer.py: confirmed an invalid frame writes blank fields, not 0 or N.
• dashboard/app.py: smoke-tested against a fake run, serves without errors.
• extract_from_bag.py / live_logger_node.py: syntax-checked only — I don't have a ROS2 install in this sandbox to test against a real bag or live topics, so treat those two as a first draft to run against your actual node and adjust the message-type / objects_json assumptions above if they don't match.