Vision-Based Surveillance for Malpractice Detection: An Analysis of Pose Estimation and Object Detection

Abstract

This paper introduces a smart, real-time surveillance system for malpractice detection by combining advanced object detection and human pose estimation. The object detection part uses established techniques along with Convolutional Neural Networks (CNNs) to effectively identify and track objects in video streams. This method has proven highly effective, demonstrating top-tier accuracy (MOTA) and precision (MOTP) on benchmark datasets like MOT16 and MOT17. To analyze human behavior, the system employs an enhanced YOLOv8 model for pose estimation, which improves both speed and accuracy. This model features two key upgrades: a Sim DLKA attention mechanism to better focus on medium-to-large targets and a new DCIOU loss function that makes training more stable and efficient. These improvements result in a 2.7% boost in mAP (mean Average Precision) and faster frame rates on standard
datasets like COCO and MPII. This combined platform provides a robust solution for monitoring in complex environments,
suitable for applications from security and traffic control to advanced human motion analysis.