A New Approach for Abnormal Human Activities Recognition Based on ConvLSTM Architecture.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Recognizing abnormal human activities in videos is challenging due to limited datasets.
• Existing datasets often lack diverse non-standard behaviors like theft or harassment.
• Current datasets may have short event durations, impacting neural network performance.

Purpose of the Study:

• To create a comprehensive dataset for abnormal human activity recognition.
• To develop and evaluate a ConvLSTM neural network for abnormal activity detection.
• To benchmark the new dataset against various deep learning architectures.

Main Methods:

• A novel dataset was created encompassing activities like Begging, Drunkenness, Fight, Harassment, Hijack, Knife Hazard, Normal Videos, Pollution, Property Damage, Robbery, and Terrorism.
• A Convolutional Long Short-Term Memory (ConvLSTM) neural network was designed and trained using the new dataset.
• The dataset was also tested with 3D Resnet50, 3D Resnet101, and 3D Resnet152 architectures.

Main Results:

• The designed ConvLSTM network achieved 96.19% classification accuracy.
• The model demonstrated 96.50% precision in identifying abnormal human activities.
• The created dataset proved effective for training and testing deep learning models for this task.

Conclusions:

• The developed dataset and ConvLSTM model significantly advance abnormal human activity recognition.
• The findings highlight the effectiveness of the proposed dataset for training robust detection systems.
• This work provides a valuable resource for future research in video-based abnormal activity detection.