MBAZOR, GOSHEN AKACHUKWU (2026) DEVELOPMENT OF AN AI-BASED HUMAN ACTION RECOGNITION SYSTEM. Other thesis, Godfrey Okoye University, Enugu.
|
Text
DEVELOPMENT OF AN AI-BASED HUMAN ACTION RECOGNITION SYSTEM.pdf.pdf Download (2MB) | Preview |
Abstract
This project describes the design, implementation, evaluation, and deployment of ActionNet – a lightweight and real-time Human Action Recognition (HAR) system implemented without any usage of GPU hardware and deep learning. The system was implemented to solve the problem of costly GPU-powered HAR systems that can be run only on professional computers and specialized cloud-based platforms. The system was designed and implemented in Python using Streamlit, OpenCV, MediaPipe Pose, and scikit-learn libraries. The system detects 33 landmarks per frame (x, y, z, visibility), and then a 15 frames sliding window is converted into a feature vector of size 410 dimensions including means (132), standard deviations (132), ranges (132), and seven scale-normalized geometric features together with their variances (14). Four scikit-learn classification pipelines (Logistic Regression – 82.04%, Ridge Classifier – 82.04%, Random Forest – 79.84%, Gradient Boosting – 88.12%) were trained and tested on a custom data set of 7,463 pose rows of 8 action types (Arms Crossed, Bending, Clapping, Hand Raised, Jumping, Sitting, Standing, and Walking). The deployed application provides the following functionality: Video upload, live WebRTC webcam analysis, session history analysis, and confidence threshold tuning.
| Item Type: | Thesis (Other) |
|---|---|
| Subjects: | Q Science > Q Science (General) |
| Divisions: | Faculty of Engineering, Science and Mathematics > School of Electronics and Computer Science |
| Depositing User: | MICHAEL MADUBUKO |
| Date Deposited: | 22 Jul 2026 12:12 |
| Last Modified: | 22 Jul 2026 12:12 |
| URI: | http://eprints.gouni.edu.ng/id/eprint/6010 |
Actions (login required)
 |
View Item |
