Driver Risk Behavior Dataset for Embedded Vision Applications
Description
This dataset contains annotated images of simulated driver behaviors captured in real driving environments. The four labeled behavior classes are: Neutral: normal driving posture without distractions Distraction: visual deviation from the road (e.g., looking away) Microsleep: eyes closed or semi-closed, simulating drowsiness Cellphone Use: using a mobile phone while driving Images were collected using an AMB82-Mini embedded board positioned in the co-pilot seat, with a camera facing the driver. Data acquisition was conducted in multiple vehicles and lighting conditions to ensure variability and realism. All images are labeled in YOLO format and anonymized to remove any identifiable facial features. This dataset is intended for training and evaluating real-time computer vision models on low-power embedded systems for driver monitoring tasks. Format: Image format: .jpg Annotation format: YOLO (.txt) Resolution: 1920×1080 px
Files
Steps to reproduce
Download and unzip the dataset archive. All .jpeg images and their corresponding .txt annotation files are located in the same folder. Each .txt file shares the same name as its corresponding image and contains YOLO-format annotations: Format: <class_id> <x_center> <y_center> <width> <height> (normalized values) Class labels: 0 = Neutral 1 = Microsleep 2 = Distraction 3 = Cellphone Use Load the dataset into any YOLO-compatible framework (YOLOv5, YOLOv4, Darknet) and update the configuration to reflect: Number of classes = 4 Input image format = .jpeg Split the dataset into train, val, and test as needed, or apply cross-validation if preferred.