esc

Type at least two characters to start searching.

↑↓ select↵ openesc close

Share

Export Citation

APA
MLA
Chicago
Harvard
Vancouver
BIBTEX
RIS
Universitas Hasanuddin
Research output:Contribution to journal›Article›peer-review

Domain-Specific Geometric Keypoint Features for LSTM-Based Shoplifting Detection in Surveillance Videos

Irfanullah M.S.R.

2026 IEEE International Conference on Artificial Intelligence and Mechatronics Systems Aims 2026

Published: 2026

Abstract

Shoplifting remains a major challenge for retail surveillance because manual monitoring is prone to fatigue, while many automated methods rely on raw pose coordinates that are sensitive to camera position and scale. This study proposes a lightweight real-time shoplifting detection system that combines YOLO-based human pose estimation, ByteTrack tracking, domain-specific geometric feature extraction, and an LSTM network with soft attention. Instead of using all 17 raw keypoints, the system extracts 10 biomechanically motivated features from 9 upper-body keypoints, including joint angles, wrist-to-hip distances, wrist-to-nose distances, and relative wrist positions normalized by torso length. These features are modeled as 15-frame temporal sequences to classify normal and suspicious behavior while providing attention-based frame importance visualization. The complete three-model ablation reveals that the attention mechanism substantially improves recall from 79.89% to 99.46% (reducing the false negative rate from 20.11% to 0.54%) at the cost of only 2,240 additional parameters, demonstrating that attention provides a substantive contribution to detection reliability rather than interpretability alone.

Other files and links

Fingerprint

Artificial intelligenceSciences
Computer scienceSciences
Computer visionSciences
Feature (linguistics)Sciences
Pattern recognition (psychology)Sciences
Remote sensingSciences
Object detectionSciences
Field (mathematics)Sciences
Noise (video)Sciences
Feature extractionSciences