Dataset Hub

The dataset introduces 1,200 samples of nighttime infrared surveillance video behavior recognition. It uses raw, unedited footage covering 10 distinct actions, each with 120 samples, following the UCF‑101 naming convention (e.g., v_DoubleWave_g01_c01). All videos were recorded outdoors at night in realistic surveillance locations such as parking lots, gardens, and alleys, involving 15 participants of varied height and build.

The dataset comprises 38 series of 30‑view RGB‑D video sequences, each accompanied by camera parameters, foreground masks, SMPL models, and assorted point‑cloud and mesh files. Every video is captured at 4K resolution, 25 FPS, and lasts between 1 and 19 seconds. All 30 viewpoints were recorded using Azure Kinect devices in a unified surrounding scene.