| Title |
Graph Transformer SlowFast Model-based Child Abnormal Behavior Analysis System |
| Authors |
윤창섭(Chang-seob Yun) ; 박상현(Sang-hyun Park) ; 박윤하(Yun-ha Park) |
| DOI |
https://doi.org/10.5573/ieie.2023.60.4.71 |
| Keywords |
Child behavior analysis; Real-time; CCTV; SlowFast; Graph transformer |
| Abstract |
This paper proposes a Graph Transformer SlowFast model capable of analyzing 4 types of abnormal child behavior to prevent child safety accidents in advance. Previous child behavior analysis models transform the pose estimation model into a 3D coordinate system and use a classification technique based on the RRT(Rapidly Explore Random Tree) algorithm. However, when RRT initial 3D coordinate inputs are insufficient, we cannot obtain optimal classification results. Also, the pose estimation and classification model will be processed as a single stream, resulting in delays in simultaneous analysis of multiple people. The proposed model separates spatial and temporal features in order to enable the training of many feature information, and, for the association between those features, the Transformer model is changed to operate on graph basis. This training of separated features is processed in parallel by creating two GPU-based streams. Regarding the analysis of children's abnormal behavior, the proposed model shows a 10% improvement of accuracy based on F1-Score, and 7FPS improvement in analysis speed over the previous models. |