Extracted track features and trained model weights for: Percentile-Calibrated Track-Level Two-Stream One-Class Framework for Human Action Anomaly Detection in Surveillance Video

This archive contains the extracted per-track feature representations and thetrained model weights underlying the anomaly-detection experiments reported inthe accompanying article, deposited separately from the journal supplementarymaterial because of their size. Contents: - features/ — per-video NPZ files holding the fixed-length track-window descriptors used in the study. Each window is 16 frames with a stride of 8 within a single track, and each descriptor is 571-dimensional: a 512-d appearance embedding from a frozen ImageNet-pretrained ResNet-18 applied to the expanded track region, followed by a stored 512-to-512 linear projection, concatenated with a 59-d motion descriptor combining 17 pose keypoints normalized to the track box with eight bounding-box kinematic features. - weights/ — the trained LSTM autoencoder checkpoints for both streams and all three random seeds, the per-stream standardization statistics, and the fixed feature projection. SHA-256 checksums for every weight file are listed in MANIFEST.md. - code/ — the feature-extraction, calibration and evaluation code. - README.md — file inventory and two reproduction paths: recomputing the reported values from the released features and weights, and rebuilding the features from source video. The source video recordings are not included. They contain identifiablefootage of volunteer participants and cannot be redistributed publicly; theyare available from the corresponding author on reasonable request, subject tothe informed-consent terms under which they were collected. The evaluation tables, raw per-frame predictions, split configuration,annotation intervals and a verification script that recomputes every reportedvalue are provided as supplementary material with the article itself.