Raw Video Data
(Video Clips with Labels)
Load and Organize
Dataset Preparation
(Load Metadata, Binary Labeling,
Stratified Splitting into Train/Valid/Test)
Split Datasets
Preprocessing
(Frame Extraction, Resizing to 224x224,
RGB Conversion, Padding, Normalization)
Process Frames (for Training)
Data Augmentation
(Horizontal Flip, Rotation, Random Crop,
Brightness/Contrast, Gaussian Blur)
(Applied to Training Data Only)
Feed Augmented Data
Feed Non Augmented Data (for Validation/Test)
Modular Model Architectures
(VGG16-RNN, C3D-style 3D CNN,
MobileNetV2-LSTM, DenseNet121-GRU,
I3D-style 3D CNN)
Train Models
Training Procedures
(Adam Optimizer, Binary Cross-Entropy Loss,
Class Weighting for Imbalance,
Early Stopping, Learning Rate Reduction)
Assess on Test Set
Evaluation
(Metrics: Accuracy, AUC, Precision, Recall, F1-Score,
Visualizations: Confusion Matrix, Precision Recall Curve)
Generate Results
Classification Output
(Normal or Suspicious Activity,
Model Performance Insights)
by Mushi
