Activity Detection using Ensemble Learning Model Trained on Customized Dataset

Abstract

Unethical activities during examination is a task of examiners; however, this work can be automated using AI-based systems to increase the quality of examinations. Many possible ways are adopted by some students during examinations to cheat and pass the exam, like using any material, by asking friends, using hand or eye movements, and in any other ways. This issue can be addressed if a dataset contains such data. In this work, these issues are addressed by developing a dataset and by designing a model for automatic malpractice detection. A dataset is prepared containing 7226 images from 4 different classes, like cheating, giving code, looking at a friend, and normal. The images are taken in different lighting conditions to make the training robust to work in real-time accurately. The dataset is used to train the different forms of YOLO models, like YOLOv8, YOLOv11, and YOLOv26 in a fashion of ensemble learning. Common detection methods often give overlapping bounding boxes. This problem is handled using majority voting, Weighted Boxes Fusion (WBF), and the non-maximum suppression (NMS) technique to keep only the bounding box with maximum confidence. The individual models like YOLOv8, v11, and v26 provided 85%, 88.3%, and 91.7% accuracy, respectively, whereas the overall accuracy increased to 93% with the ensemble technique.