An Efficient Lightweight CNN Autoencoder for Skeleton-Based Video Anomaly Detection

Received: 24 March 2026 | Revised: 26 April 2026 | Accepted: 1 May 2026 | Online: 6 June 2026

Corresponding author: Mostafa Ibrahim Labib

Abstract

Anomaly detection based on human motion has been an active area of research in computer vision, evolving from traditional handcrafted methods to deep learning–based representations. This paper proposes a lightweight Convolutional Neural Network (CNN) autoencoder designed for anomaly detection using skeleton image representations. Unlike prior Long Short-Term Memory (LSTM)-based approaches that depend on temporal coordinate sequences, the proposed approach encodes skeletal joint relations as 2D spatial maps, enabling convolutional learning of local and global structure. The model was trained exclusively on normal samples from structured datasets generated from pose-estimation outputs to reconstruct typical poses, with reconstruction error used as an anomaly score. Skeleton images are preprocessed before training using grayscale conversion, Gaussian blurring, and resizing to ensure consistency, reduce noise, and improve efficiency. Experimental results achieved an accuracy of 59.6%, precision of 61.0%, and recall of 95.5%, yielding an overall F1-score of 74.5%. These results demonstrate that the CNN-based autoencoder successfully identifies most anomalous poses while maintaining reasonable precision, validating the benefits of spatial learning and augmentation for skeleton image anomaly detection.