Work

Framework for real-time bounding box detection of Indian dance movements for multiple persons on a 5G multi-access edge computing server

Year: 2026
Type: article
Abstract: This paper presents a machine learning workflow, run on a 5G multi-access edge computing (MEC) server, for detecting and classifying the Kuchipudi dance movements of multiple dancers. Dance movements in three Kuchipudi videos were annotated with the Computer Vision Annotation Tool (CVAT), and a YOLO11s model was trained to locate them with bounding boxes. For classification, the annotated regions were cropped, downscaled to reduce computation, lowpass filtered using the two-dimensional discrete Fourier transform, and enhanced with contrast limited adaptive histogram equalisation (CLAHE). Because the classes were highly imbalanced, StyleGAN2-ADA was used to generate synthetic images, giving 1500 images in each of seven classes. Two- and three-layer convolutional neural networks (CNNs) were compared with ResNet18 and ResNet50. ResNet50 performed best, with a validation accuracy of 89% and a macro F1-score of 84%. On the CPU-only MEC server, inference took 44.68 ms per frame on average, and the end-to-end throughput was 2 frames per second.
Language: English
Open Access status: diamond
OpenAlex ID: W7236723852
FWCI: 1.056
Cites: 36
Cited by:
Related to: 10