UNISV‑Dataset Overview

Dataset Description

• Sample Count: 1,200 samples.
• Action Classes: 10 different behaviors, each with 120 samples.
• Video Characteristics: Raw or randomly edited original videos, most without people. Filenames follow the UCF‑101 format, e.g., "v_DoubleWave_g01_c01" where "v_DoubleWave" denotes the action class, and "g01" and "c01" indicate location and sample index.
• Scene Setup: All videos were recorded outdoors at night, mirroring key positions of real surveillance cameras such as parking lots, gardens, and alleys.
• Participants: 15 volunteers with diverse heights and body types.

Action Classes

• Class List: "Double Wave," "Wave Hand," "Walk," "Jump," "Squat," "Jogging," "Push People," "Shake Hands," "Embrace," and "Fight."
• Participant Count: Each class involves 6 participants, totaling 15 participants overall.
• Video Samples per Class: 120 videos per class, 1,200 videos in total.
• Video Specs: 10 fps, resolution 480 × 248 px.

Citation Information

If you use this dataset, please cite:

Feng Z, Wang X, Zhou J, et al. MDJ: A Multi‑Scale Difference Joint Keyframe Extraction Algorithm for Infrared Surveillance Video Action Recognition[J]. Digital Signal Processing, 2024: 104469. https://doi.org/10.1016/j.dsp.2024.104469.