Postures anomaly tracking and prediction learning model over crowd data analytics.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Modern advancements in intelligent systems necessitate effective crowd analysis, particularly in e-learning environments.
• Observing and analyzing crowd behavior, including normal and abnormal actions, presents significant challenges.
• Existing methods struggle with complex crowd data, requiring improved tracking and prediction frameworks.

Purpose of the Study:

• To propose an organized method for multiobject tracking and action prediction in e-learning crowd data.
• To develop a framework capable of distinguishing between normal and abnormal activities using multilayer perceptron.
• To enhance the accuracy and efficiency of crowd behavior analysis in educational technology.

Main Methods:

• Feature extraction using fused dense optical flow, gradient patches, super pixel, and fuzzy c-mean.
• Multiobject tracking implemented with compressive tracking and Taylor series predictive tracking.
• Data complexity reduction via T-distributed stochastic neighbor embedding (t-SNE) and classification using multilayer perceptron (MLP).

Main Results:

• The proposed framework achieved a mean accuracy of 87.00% across three diverse crowd activity datasets.
• Specific dataset accuracies include 85.75% for USCD-Ped and 88.00% for the IITB corridor dataset.
• The method effectively extracts trajectories and predicts actions, demonstrating robust performance.

Conclusions:

• The developed e-learning crowd analysis framework demonstrates high accuracy in predicting normal and abnormal actions.
• The integration of advanced feature extraction and tracking algorithms significantly improves crowd behavior analysis.
• This research offers a valuable tool for enhancing safety and understanding in e-learning environments through intelligent crowd monitoring.