Research Projects
Predicting Near-Miss Accidents Using Dashboard Camera Data
Warns of a crash 1–5 seconds before it happens using only a cheap dashcam, and shows why. Caught 84% of 900 real near-miss videos in time.
- Role
- Principal Investigator
- Context
- Student Scientific Research Project, VNUHCM University of Information Technology
- Supervisor: Dr. Dao Minh Son, Faculty of Computer Science
- Period
- Aug 2023 – Feb 2024
Overview
A lightweight, explainable pipeline that predicts collision points 1–5 seconds ahead using only a single dashcam, designed for embedded or resource-constrained devices as a low-cost ADAS alternative.
Converts the driver's first-person view into a top-down metric view, then builds object trajectories to explain how a potential accident would occur (Explainable AI), rather than relying on a black-box deep model.
Key achievements
Evaluated on 900 near-miss videos from the Hyjari dataset (Tokyo University of Agriculture and Technology): 84% of cases warned 1–5 seconds before the near-miss event, 12% missed, and 4% warned late.
Responsibilities
- Led the project end-to-end as the sole principal investigator: literature review, pipeline design, experiments, and final report.
- Designed camera calibration based on the pinhole model: built the intrinsic matrix from dashcam specs (focal length, FOV) and estimated extrinsic rotation from vanishing points.
- Fine-tuned DeepHough (Hough transform + CNN-FPN) on 1,000 street images from the Hyjari dataset to detect semantic lines, and designed an intersection-clustering algorithm to locate the vanishing point per frame.
- Added motion-vector-based ego-car turn detection to estimate the vanishing point when lines are unavailable during left/right turns.
- Combined YOLOv7 and DeepSORT for vehicle detection and tracking, converting pixel coordinates into vehicle-centric world coordinates (meters).
- Computed a dynamic safe region using the AASHTO stopping-distance formula and predicted collision points with LSTM / Linear Regression over tracked trajectories.