A Lightweight Digital-Twin-Based Framework for Edge-Assisted Vehicle Tracking and Collision Prediction

Vehicle tracking, motion estimation, and collision prediction are fundamental components of traffic safety and management in Intelligent Transportation Systems (ITS). Many recent approaches rely on computationally intensive prediction models, which limits their practical deployment on resource-constrained edge devices. This paper presents a lightweight digital-twin-based framework for vehicle tracking and spatiotemporal collision prediction that relies solely on object detection, without requiring complex trajectory prediction networks. The framework is implemented and evaluated in Quanser Interactive Labs (QLabs), a high-fidelity digital twin of an urban traffic environment that enables controlled and repeatable scenario generation. A YOLO-based detector is deployed on simulated edge cameras to localize vehicles and extract frame-level centroid trajectories. Offline path maps are constructed from multiple traversals and indexed using K-D trees to support efficient online association between detected vehicles and road segments. During runtime, consistent vehicle identifiers are maintained, vehicle speed and direction are estimated from the temporal evolution of path indices, and future positions are predicted accordingly. Potential collisions are identified by analyzing both spatial proximity and temporal overlap of predicted future trajectories. Our experimental results across diverse simulated urban scenarios show that the proposed framework predicts approximately 88% of collision events prior to occurrence while maintaining low computational overhead suitable for edge deployment. Rather than introducing a computationally intensive prediction model, this work introduces a lightweight digital-twin-based solution for vehicle tracking and collision prediction, tailored for real-time edge deployment in ITS.

💡 Research Summary

The paper introduces a lightweight, edge‑oriented framework for vehicle tracking and collision prediction that avoids the heavy temporal deep‑learning models commonly used in intelligent transportation systems (ITS). Instead, it relies on a single‑stage object detector (YOLOv11) running on simulated edge cameras and on a high‑fidelity digital‑twin environment (Quanser Interactive Labs, QLabs) to generate offline path maps. The methodology consists of four stages.

1. Data Collection & Offline Path Generation – Multiple traffic scenarios are executed in QLabs, each containing a vehicle that follows a predefined route. YOLOv11 processes every video frame, extracting bounding‑box centroids. All centroids belonging to the same logical route are aggregated, then resampled to a fixed spatial resolution, yielding an ordered pixel‑wise representation (R_p) for each path (p). These path files are stored for later use.

2. Online Path Association via K‑D Trees – For each incoming detection, the framework builds a K‑D tree for every stored path (once, offline) and queries the nearest point to the detection centroid. If the Euclidean distance is below a configurable threshold (D_{path}), the detection is associated with that path and its index (i_p) is recorded. The logarithmic search complexity ((O(\log N))) dramatically reduces the computational load compared with a naïve linear scan, making the operation suitable for low‑power edge processors.

3. Vehicle Tracking & Future Path Estimation – A simple distance‑based data association algorithm assigns persistent IDs to detections across frames. When a detection lies within a tracking radius (D_{trk}) of an existing track, it is linked; otherwise a new track is created. The resulting trajectories (T_v) are linked to path indices, forming a history (H_v). By examining the change of path index over time (with optional temporal down‑sampling parameters (L) and (K)), the system estimates each vehicle’s average speed and heading direction. Future positions are predicted by linearly extrapolating the current index forward for a configurable horizon (N) (number of future points).

4. Spatio‑Temporal Collision Prediction – For every pair of vehicles, the framework generates predicted future trajectories on their respective paths. A collision is flagged when two trajectories come within a spatial distance threshold (D) while their predicted arrival times differ by less than a temporal tolerance (\Delta t). All possible path‑pair combinations ((N_{comb})) are examined; the number of combinations that satisfy the collision condition ((N_{col})) yields an estimated collision probability (Pr_{col}). The final output is a collision summary set (S) containing the most likely collision events.

Experimental Evaluation – The authors evaluated the system in a variety of simulated urban scenarios (intersections, roundabouts, pedestrian crossings). YOLOv11 achieved >30 fps on edge‑class hardware, and K‑D‑tree association incurred <0.2 ms latency per detection. The complete pipeline consumed <15 % of a single CPU core and <120 MB RAM, confirming its suitability for devices such as NVIDIA Jetson Nano or Raspberry Pi 4. Collision prediction accuracy reached approximately 88 % with a false‑positive rate around 7 %, demonstrating that a simple linear extrapolation, when anchored to accurate path maps, can be highly effective.

Key Contributions

• Leveraging a high‑fidelity digital twin (QLabs) to generate reproducible, scenario‑rich training data without safety risks.
• Proposing a YOLO‑based, K‑D‑tree‑accelerated path association mechanism that eliminates the need for heavyweight tracking filters (e.g., Kalman, particle filters).
• Demonstrating that persistent ID assignment combined with path‑index history suffices for reliable speed/direction estimation and short‑term trajectory prediction.
• Providing a complete, reproducible pipeline (offline path generation, online association, tracking, prediction, collision scoring) with open‑source‑style algorithmic descriptions.

Implications & Future Work – By removing deep temporal models, the framework dramatically reduces inference latency and power consumption, enabling real‑time deployment on edge nodes in smart‑city surveillance, V2I safety services, and low‑cost traffic monitoring installations. Future research directions include (i) bridging the domain gap between simulated and real video streams through domain adaptation, (ii) integrating additional sensors (LiDAR, radar) to improve robustness under adverse weather or occlusion, and (iii) extending the prediction horizon with lightweight probabilistic models (e.g., Gaussian processes) while preserving the edge‑friendly computational budget.