Suspicious Activity Detection — MediaPipe Pose Landmarks

The dataset comprises MediaPipe Pose landmarks extracted from an image dataset created for binary classification of suspicious behaviors in indoor surveillance settings. The dataset is intended for use in developing and testing a multimodal surveillance system that combines visual appearance features generated from EfficientNet-B0 and structural pose features analyzed using the XGBoost classifier.

Data CollectionImages were collected with a webcam using a real-time overlay of MediaPipe Pose landmarks. Data was recorded with four subjects performing suspicious and regular behavior over multiple recording sessions in different lighting (daylight, artificial, low-light), attire, and camera angles in an indoor setting. Suspicious behavior included actions such as crouching, reaching, bending, and concealing. Regular behavior included standing, walking, and sitting.

Pose landmarks were extracted from each image using MediaPipe Pose (version 0.10.14) with the following configuration:

static_image_mode = True

model_complexity = 1

min_detection_confidence = 0.5

For each image, all 33 body landmark (x,y) coordinate pairs were extracted and concatenated into a 66-dimensional feature vector. Landmarks with a visibility score below 0.5 were zero-padded to preserve consistent vector dimensionality. Images for which pose extraction failed entirely (57 of 1,140 total images) are excluded from this dataset.