SentinelCAR Dataset

The SentinelCAR Dataset addresses the growing need for security in the context of electric mobility and shared vehicles. The dataset consists of annotated images captured with a single fisheye camera mounted on the vehicle ceiling, providing a 180° view through the four windows. To preserve privacy and focus only on relevant activity, the images include applied masks, so that only the content visible through the windows is available for analysis. The dataset supports research on object detection and activity classification, and accompanies the study “Sentinel Mode Using Sensor Fusion for Improving Security in Shared Vehicles” (published in IEEE Access).

Contents:

- Images: 18,308 frames in JPG/PNG format, captured from shared vehicle scenarios.

- YOLO Annotations: 14,108 text files in YOLO format, with two activity-based classes:

0 – Person_Active: Person actively interacting inside the vehicle.

1 – Person_Not_Active: Person present but not actively engaged.

Classification Strategy:

In addition to bounding-box annotations, an image-level classification dataset is derived automatically:

Active: Assigned if at least one Person_Active (class 0) is present.

Not_Active: Assigned if the image contains only Person_Not_Active (class 1) or no persons at all.

This straightforward yet effective approach enables image-level activity classification directly from the object detection annotations.

Included Scripts:

- prepare_classification_dataset.py – Generates the classification dataset using the above logic. Splits data into train/val/test sets while maintaining class balance and directory structure.

- split_object_detection.py – Organizes the dataset for YOLO object detection training, preserving the correspondence between images and bounding-box annotations across train/val/test splits.

Potential Applications:

- Training models for detecting Active vs. Not_Active persons in shared vehicles.

- Research in computer vision for automotive security and surveillance.

- Benchmarking supervised learning methods for object detection and activity classification in real-world mobility scenarios.