Input and output contracts
Paths below describe relative file locations and examples. Use command-line arguments to specify input and output files. Original inputs are read only; existing output files are not overwritten. Keep the image view used for pixel centers and bounding-box corners identical to the view used for their calibration.
| Stage | Existing input to supply | Output and next consumer |
|---|---|---|
| Coordinate conversion | One named data/raw/SCENE_raw_detection_data_N.csv, actual scene JSON and matching videoN |
Original fields plus Source_ID,x_center,y_center in meters. Input remains one observed frame/vehicle per row. |
| Lane join | Metric output plus a source-frame annotation CSV containing Frame,Vehicle_ID,Lane number,Direction |
Metric rows with Lane number,Direction, joined one-to-one using the original source Frame,Vehicle_ID keys. |
| Reconstruction | Metric centers with lanes/direction, actual FPS and explicit offsets | One row for every reconstructed source frame/vehicle, plus separate rejected rows and per-vehicle screening measurements. |
| Length estimation | Original six-column source detections and same scene/source calibration | One row per original vehicle: fixed length, bbox lateral diagnostic, frame counts and quality. |
| Length joining | Reconstructed trajectories with lineage and a length table; otherwise an existing verified ID mapping | Basic trajectories with one fixed Vehicle_Length_m per final vehicle. QC fields stay in the separate length table. |
| Risk | Released Basic with valid lengths, lanes, direction and Time_s or Delta_t_s | Frame-wise TTC/DRAC, event-frame PET, risk target and risk level. Delta_t_s input additionally receives Time_s and Time_Reference_ID. |
| Pair generation | One released observation-group Basic CSV, or one source-labelled reconstructed Basic CSV, with meter positions and velocities | Frame-sorted directed candidate pairs and a JSON runtime summary. |
| Balanced fitting samples | Full frame-sorted candidate pairs | Frame,EV_Vehicle_ID,POV_Vehicle_ID,x_local,y_local,v_rel,cangle,round_v,Sample_Group_ID; the group summary records sample counts and discarded remainders. The fitting subset contains the listed fields. |
| Parameter inference | Balanced fitting subset; default minimum and maximum used per category/speed key are 50000 | Parameter CSV with one accepted result per category/speed group; a separate JSON records group status and sample counts. The run stops if a requested group fails its acceptance checks. |
| Interaction annotation | Original Basic, full candidate pairs and completed scenario/source-appropriate PR parameter CSV | Basic fields plus pair-relative states, PR, interaction flag and type; a repeated ego/frame can represent several targets. |
Required fields
- Raw detections:
Frame,Vehicle_ID,x_center_px,y_center_px,width_px,height_px. Integers for Frame and ID; pixel units for centers and box dimensions; unique source-local(Frame,Vehicle_ID). Each manifest row binds this local key to its scene/source. - Lane annotation:
Frame,Vehicle_ID,Lane number,Direction. Keys must refer to the original source frames/IDs before reconstruction; one record per observed key, complete coverage; direction isLeftorRight. Optional Source_ID must agree with the metric source. Apply the corresponding identifier mapping when transferring processed Basic annotations to the original detection keys. - Reconstruction output lineage:
Source_ID,Original_Vehicle_ID,Original_Frame,Frame,Ego_Vehicle_ID.Time_sis seconds from the input recording origin plus explicit time offset;Delta_t_sis blank at each track start and 1/FPS thereafter. The example commands use zero offsets and therefore retain a source-local reference. - Kinematics:
x_center,y_centerm;v_x,v_ym/s;a_x,a_ym/s². The reconstruction outputs smoothed positions under x_center/y_center. Positions use road-aligned local metric coordinates obtained by master-image registration and measured scale conversion. - Length output:
Source_ID,Vehicle_ID,Vehicle_Length_m,BBox_Lateral_Extent_Q25_m,Input_Frames,Valid_Before_IQR,Retained_Frames,Removed_By_IQR,Quality. Vehicle_Length_m is missing if fewer than five frames survive. Quality is good at >=20 retained frames, limited at 5–19, insufficient otherwise. - Optional processed-ID mapping:
Ego_Vehicle_ID,Source_ID,Original_Vehicle_ID; both final ID and source/original identity map one-to-one. Compress frame-wise mapping rows only after verifying consistency. - Required pair fields:
Frame,EV_Vehicle_ID,POV_Vehicle_ID,x_local,y_local,v_rel,cangle. x_local/y_local are local X/Y components in meters, v_rel is a positive scalar in m/s, and cangle is binary, not degrees. The generator also writes signedcangle_deg,Distance_m, andAngle_Definedfields. - PR parameters:
cangle,round_v,rx_plus,rx_minus,ry_plus,ry_minus,bx_plus,bx_minus,by_plus,by_minus. Category/speed key unique; all eight parameters positive. Additional fit-status fields are allowed.
Calibration fields and source matching
Each measured JSON includes master_source, the two-coordinate origin_px, rotation_rad=0, two measured reference_segments, and sources. Each non-master source contains matched Nx2 fitting arrays src_points and master_points. The master source defines the reference view and uses the identity source-to-master mapping. Use the measured files in calibrations/ for the supplied recordings. Field meanings and units are detailed in CALIBRATION.md.
The sixteen filename/source mappings and frame rates are listed in configs/source_manifest.csv. Process one source per run. For combined sources, provide established offsets with --id-offset, --frame-offset, and --time-offset. Use source-frame annotation keys for the lane join; apply any final-ID mapping only at the appropriate later stage. Measured --merge-x-m and --diverge-x-m values must use the same metric frame as the trajectories.
Error behavior and outputs
Invalid numeric values, duplicate identity keys, source disagreements and missing required fields fail explicitly. Invalid frame-level bounding boxes are excluded and counted by the length estimator. A length table can contain a vehicle with a missing estimate, but a missing vehicle mapping fails. The main risk example uses the existing per-vehicle lengths. Fitting samples use relative speeds up to 30 m/s and complete sparse blocks; runtime summaries record excluded observations and block remainders.
All supplied candidate pairs remain eligible for final annotation, including those excluded from fitting by the speed cutoff. Parameter lookup selects the nearest fitted relative-speed bin within each angle category, with no default maximum speed difference, including at the fitted-range boundaries. Set --speed-tolerance to require a bin within a specified speed difference; an unmatched lookup raises an error.
Released Basic input and time references
The primary downstream input is the released observation-group Basic CSV, with Ego_Vehicle_ID,Frame,Delta_t_s,x_center,y_center,v_x,v_y,a_x,a_y,Direction,Lane number,Dist_to_Merge_m,Dist_to_Diverge_m,Vehicle_Length_m. This entry point uses the lane and direction fields already present in the Basic file. The pair generator uses metric centers and velocities directly; annotation uses the existing processed IDs, lane and direction fields.
For risk input without Time_s, positive Delta_t_s values define the elapsed time into the corresponding processed frame. Values for the same frame must agree. Each vehicle must have consecutive frames and a positive interval after its first record. The time builder accumulates these intervals across connected frames. A missing incoming interval or a gap in observed frames starts a new time reference at zero. Time_Reference_ID identifies that reference, and each vehicle must belong to one reference. This represents local sequence time without assigning an unobserved interval between recordings. Existing Time_s input retains its supplied reference; explicit --fps retains the constant-rate option.
The Basic files use processed Ego_Vehicle_ID,Frame; raw detections use source-local Vehicle_ID,Frame. A source-frame annotation CSV for reconstruction is keyed by the original detection records. Transferring Basic labels uses the corresponding source/frame/vehicle mapping.