Anomaly-Detector

Anomaly Detection in Egocentric Traffic Videos using Deep Learning

---

Project Overview

This project explores two aproaches in detecting anamolies in ego-centric traffic videos

1. Unsupervised Approach: Convolutional Autoencoder (ConvAE) - reconstruction-based anomaly detection
2. Supervised Approach: VideoMAE + Classifier - fine-tuned transformer for binary classification

The models are trained and evaluated on the DoTA dataset for egocentric driving scenarios.

Prerequisites

• Python 3.8+
• CUDA-capable GPU (recommended)
• 16GB+ RAM

Installations

1. Clone the repository:

git clone https://github.com/MostafaK2/Anamoly-Detector.git
cd Anamoly-Detector

2. Install dependencies:

pip install -r requirements.txt

Required Dependencies

torch>=2.0.0
torchvision>=0.15.0
transformers>=4.30.0
opencv-python
Pillow
numpy
matplotlib
scikit-learn
tqdm

Dataset

Download DoTA dataset

• Find dataset in this drive folder https://drive.google.com/drive/folders/1_WzhwZC2NIpzZIpX7YCvapq66rtBc67n
• Command to download gdown https://drive.google.com/uc?id=FILE_ID
• Command to export 7z x DoTA_seg.7z

Generate text files

• First generaate .txt file for seperating the train/validation/test using functions split_by_json_file, and scan_and_export_json_files from utils.py
• Currently filenames already uploaded into the repository (generate if you would like)

For Videomae + Classifier

1. Update paths in the scripts to point to your dataset location and output directory for generated clips
2. Preprocess the dataset (for faster training):

python scripts_preprocessing.py

3. Currently commented out

This will create preprocessed clips in:

• preprocessed_clips/train/
• preprocessed_clips/validation/
• preprocessed_clips/test/

Training

1. Training ConvAE (Unsupervised)

cd ConvAE
python ConvAE_pipeline.py

Configuration options (edit in Config class):

• BATCH: Batch size (default: 32)
• EPOCH: Number of epochs (default: 30)
• LR: Learning rate (default: 1e-4)
• STACK_SIZE: Number of frames per sequence (default: 16)
• OVERLAP: Frame overlap between sequences (default: 8)

2. Training VideoMAE Classifier (Supervised)

cd videomae
python videomae_anamoly_pipeline.py

Configuration options (edit in Config class):

• BATCH: Batch size (default: 64)
• EPOCH: Number of epochs (default: 6)
• LR: Learning rate (default: 6.2e-4)
• FREEZE_LAYERS: Number of frozen encoder layers (default: 9)
• DROPOUT: Dropout rate (default: 0.5)
• EARLY_STOPPING_PATIENCE: Early stopping patience (default: 2)

Important paths to update:

• ROOT: Dataset root directory
• PREPROCESSED_TRAIN: Path to preprocessed training data
• PREPROCESSED_VALID: Path to preprocessed validation data

Evaluation & Visualization

Evaluate ConvAE and VideoMAE + classifier

cd ConvAE
python evaluateConvAE.py
python visualizeConvAE.py
python testVideoMAEAnamoly.py
python _videomae_visualize_results.py

Note: These codes will use configurations from ConvAE_pipeline.py and videomae_anamoly_pipeline.py code

Outputs:

• AUC-ROC score
• Reconstruction error plots
• Anomaly score distributions
• Visualization Directory ** Important paths to update **
• ROOT: Path where the actual raw dataset is located
• MODEL_PATH: Pretrained model path
• 'TEST_TXT_PATH': File containing JSON filenames for test videos
• OUTPUT_DIR: Specify where you would like to output ur evaluation or visualization results

These would create

• Evaluation results text files for both convae and videomae+classifier
• (Conv-AE) Reconstruction error plots
• (video-mae) TP, FP, FN, TN examples with frames
• (video-mae) Score distributions over the computed sequences with frames
• (video-mae) Anomaly scores over time with ground truth

Results and Visualization

MODEL Performance

Model	AUC-ROC
ConvAE	0.556
VideoMAE	0.828

Videomae+classifier further evaluation

Class	Precision	Recall	F1-Score	Support
Normal	0.7437	0.8573	0.7965	1472
Anomaly	0.8094	0.6722	0.7345	1327
Weighted Avg	0.775	0.769	0.767	2799

Acknowledgments

• VideoMAE pretrained model from MCG-NJU
• DoTA Dataset for traffic anomaly detection
• Hugging Face Transformers library