Qpu523/VPSCI-Dataset

🚦 VPSCI Dataset – Vehicle‑Pedestrian Safety‑Critical Interaction Dataset

What it is – A large‑scale, high‑resolution dataset of near‑miss vehicle‑pedestrian interactions at urban intersections. The data are generated with a three‑stage reinforcement‑learning pipeline (MA‑SST‑DDPG) that is first pre‑trained on real‑world near‑miss recordings, then refined online inside the CARLA driving simulator, and finally used to synthesize > 198 k interaction episodes.

Why it matters – Autonomous‑vehicle research needs realistic, safety‑critical scenarios to train and evaluate perception, prediction, and planning modules. Existing public datasets contain relatively few close‑call events; VPSCI fills that gap with thousands of diverse, high‑frequency trajectories that have been shown (via a human Turing‑test) to be indistinguishable from real‑world video.

---

📚 Core contents

Item	Details
Scenarios	Two intersection layouts (A & B) in CARLA Town10, each with four right‑turn vehicle / crossing‑pedestrian directions (8 total).
Episodes	>10 000 episodes per CSV file (≈ 2.6 km pedestrian path, 1.7 km vehicle path), recorded at 20 Hz.
Fields	count, frame, veh_id, ped_id, position_x_av, velocity_y_ped, CurvTTC … (full list in the accompanying paper).
Size	~198 000 interaction episodes across all eight scenario files.
Videos	144 example videos illustrating the generated interactions (download link provided).
Human evaluation	51 participants rated realism; generated videos were statistically indistinguishable from real‑world recordings (p = 0.93).

---

🛠️ How the data are created

1. Stage 1 – Pre‑training – Multi‑agent state‑space Transformer‑enhanced DDPG (MA‑SST‑DDPG) agents learn evasive behaviours from a curated set of real near‑miss events.
2. Stage 2 – Online learning – The agents continue to adapt inside CARLA, encountering a wide variety of high‑risk situations and narrowing the simulation‑reality gap.
3. Stage 3 – Large‑scale generation – The refined agents drive the simulated intersections, automatically logging every interaction episode.

---

📥 Getting the data

Subset	Direction of pedestrian crossing	Download link
1	South → North	[OneDrive link]
2	North → South	[OneDrive link]
3	East → West	[OneDrive link]
4	West → East	[OneDrive link]
5‑8	Same four directions for the second intersection layout	[OneDrive links]

(Each link points to a CSV file containing the episodes for that subset.)

---

🎬 Example videos

A folder of 144 videos is available for download to see the interactions in action: [Download video collection].

---

📊 Evaluation (Turing‑test)

Compared groups	t‑statistic	p‑value	KS‑statistic	p‑value	Significant?
CARLA baseline vs. MA‑SST‑DDPG	-13.5	<0.0001	0.4951	<0.0001	Yes
CARLA baseline vs. Real world	-13.63	<0.0001	0.5294	<0.0001	Yes
MA‑SST‑DDPG vs. Real world	-0.09	0.9282	0.0588	0.8732	No

The lack of a significant difference between the generated and real videos confirms the high fidelity of the dataset.

---

📖 Citation

If you use the dataset or the underlying generation method, please cite:

@article{pu2027generating,
  title   = {Generating realistic safety--critical scenarios for vehicle--pedestrian interactions},
  author  = {Pu, Qingwen and Xie, Kun and Zhu, Yuan and Zhai, Guocong},
  journal = {Transportation Research Part C: Emerging Technologies},
  volume  = {194},
  pages   = {106002},
  year    = {2027},
  doi     = {10.1016/j.trc.2026.106002},
  url     = {https://www.sciencedirect.com/science/article/pii/S0968090X26004882}
}

---

🙋‍♂️ Contact

• Kun Xie – kxie@odu.edu
• Qingwen Pu – qpu001@odu.edu

---

⭐ Star history

[Star history chart] (auto‑generated by the repository).

---

Bottom line: VPSCI provides a ready‑to‑use, simulation‑generated but realism‑validated collection of vehicle‑pedestrian near‑miss trajectories, ideal for training and benchmarking perception, prediction, and planning models in autonomous‑driving research.