bikeped Conformance Scenarios

24 scripted pedestrian–cyclist intersection scenarios used to evaluate the collision-warning system in bikeped. Each scenario defines a handful of agents (pedestrians, cyclists, e-bikes) moving on the ground plane through a hand-authored sequence of waypoints; the testbench interpolates these at 30 fps and assigns per-frame ground-truth danger labels from a conservative kinematic safety model.

Companion to the technical report A Real-Time Bike–Pedestrian Safety System with Wide-Angle Perception and Evaluation Testbed for Urban Intersections.

Dataset structure

Two configs:

scenarios — 24 rows

One row per scenario with the raw waypoint paths.

column	type	description
scenario_id	int	0-based index
name	string	scenario title (e.g. Bike Approaches Ped)
expected_dominant	string	ground-truth dominant label: idle / safe / warning / alert
vehicle	string	bicycle or ebike (selects the braking profile)
duration_s	float	scenario length in seconds
n_person / n_bike	int	agent counts
agents	list[dict]	each with type, agent_id, path
agents[].path	list[dict]	waypoints: x and y in metres, t in seconds
sim_video	string?	relative path to videos/sim/NN_slug.mp4 (if shipped)
carla_video	string?	relative path to videos/carla/NN_slug.mp4 (if shipped)

frames — 12901 rows

Denormalized: one row per (scenario, agent, frame) at 30 fps, with linearly-interpolated ground-plane position and the scenario-level ground-truth danger assessment for that frame.

column	type	description
scenario_id	int
scenario_name	string
frame_idx	int	0 … ⌊duration·fps⌋
time_s	float	frame_idx / fps
agent_id	int
agent_type	string	person or bike
x_m, y_m	float	interpolated world position
is_danger	bool	physically dangerous (collision trajectory)
actionable	bool	is_danger AND warning can still help the ped
severity	float	0–1, kinetic-energy proxy
ttc_s	float?	time to collision (s); null = no collision
closing_speed_ms	float	m/s; positive = approaching
bike_speed_ms	float	m/s
stopping_dist_m	float	required stopping distance at current speed
warning_budget_s	float?	seconds of warning before closest approach
distance_m	float?	current nearest bike–ped distance

Videos

Two renderings of each scenario, shipped alongside the JSONL:

• videos/sim/NN_slug.mp4 (24/24) — schematic bird's-eye-view rendered by sim_visualizer.py. Ground-plane plot with crosswalk overlay, agent markers, trajectory traces, and the decision pipeline state (SAFE/WARN/ALERT/IDLE).
• videos/carla/NN_slug.mp4 (24/24) — photorealistic pole-mounted camera view rendered in CARLA 0.9.16 (Town10HD_Opt) by carla_scenario.py. Scenario agents are spawned as CARLA pedestrians and cyclists along the same waypoint paths; the camera captures RGB frames from the pole location (x=5, y=-13, z=3.66) at 30 fps.

Both videos use the same scenario numbering (00 … 23) as scenario_id.

Coordinate frame

World coordinates in metres. The crosswalk runs along y at x = 5 (≈3 m wide). The deployed camera sits at (x=5, y=-13, z=3.66) looking across the crosswalk. Agent positions are world-frame; in the live system the fisheye inverse-projection produces camera-relative BEV coordinates.

Kinematic ground-truth model

A frame is labelled dangerous if the cyclist is closing, the closest point of approach is within 5 m, and either the stopping distance exceeds 80% of the remaining gap or TTC < 3 s. Conventional bicycles use t_react = 0.84 s, decel = 1.96 m/s² (85th-percentile field measurements, Martin & Bigazzi 2025, n=302). E-bike scenarios (Fast Approach, E-Scooter, Accelerating E-Bike) use decel = 6.0 m/s² (Naude et al. 2025 disc-brake track tests). Severity weights s = min(v²/v_max², 1) with v_max = 12 m/s.

Scenario categories

category	scenarios
Safe crossings	3
Standard approaches	5
High-speed / accelerating	3
Accessibility	2
Multi-agent	3
Edge cases	3
Non-linear trajectories	4

21 of the 24 scenarios contain ground-truth danger intervals.

Citation

@misc{turkcan2026realtimebikepedestriansafetywideangle,
  title         = {A Real-Time Bike-Pedestrian Safety System with Wide-Angle Perception and Evaluation Testbed for Urban Intersections},
  author        = {Mehmet Kerem Turkcan},
  year          = {2026},
  eprint        = {2604.17046},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2604.17046}
}

Code: github.com/mkturkcan/bikeped

License

CC-BY-4.0. Scenarios are hand-authored and distributed as a dataset for evaluating pedestrian–cyclist conflict warning systems.