Anomaly Dectection in Videos

Objective

Video anomaly detection system with multiple algorithms, and real-time support.

Currently Implemented Approaches

For each approach, there should be a jupyter notebook, evaluation support (taking a sample test and output whether it is anomaly or not), and real-time support.

Approach	Notebook Status	Evaluation Support	Real Time Support
STAE	todo	done	todo
LSTM Autoencoder	done	todo	todo

Configurations

Create a new Config.py by copying Config.py.example, which contains the following parameters.

• DATASET_PATH: path to USCDped1/Train directory.
• SINGLE_TEST_PATH: the test sample you want to run.
• RELOAD_DATASET: boolean parameter. set to True if when reading the database the first time or False to read from cache.
• RELOAD_TESTSET: boolean parameter. set to True if when reading the test sample the first time or False to read from cache.
• RELOAD_MODEL: set to True if you want to re-train the model, False otherwise.
• CACHE_PATH: path to the cache file.
• BATCH_SIZE & EPOCHS: parameters for training the model.
• MODEL_PATH: the path to save the model.

Evaluation

After putting the desired configurations, run evaluation.py to get the result of the chosed sample test after processed by the model.

Notes

LSTM autoencoder which I used in my article only exists as a jupyter notebook in notebooks/lstmautoencoder. It'll be integrated with the project later.