/Anomaly_detection

This repository contains all the codes related to the project 'Anomaly Detection in Surveillance Videos'. Codes related to model training, testing, summarization are present along with documentation. The project also uses an analysis method known as Grad-CAM which highlights regions of input, which played a role in the decision making of the model.

Primary LanguagePython

Anomaly_detection

This repository contains all the codes related to the project 'Anomaly Detection in Surveillance Videos'. Codes related to model training, testing, summarization are present along with documentation

Directory structure:

  1. video_root_path – must be the root directory for your datasets
  2. Create a folder corresponding to your dataset inside the video_root_path directory. Inside the folder, place the training and testing videos inside training_videos and testing_videos folders.
  3. Also create folders named training_frames and testing_frames inside the dataset folder. This is where the frame extraction program will store the extracted frames.
  4. Additionally, with the available datasets, groundtruth, reconstruction-frame plots and rating files can be found under groundtruth, graph_plot, rating_training and rating_testing folders. For new datasets, these folders should be created manually.
  5. data_preprocessing.py should be placed along with the training code in the same directory.

Training:

  1. unsupervised_training.py Command: python3 unsupervised_training.py Code for training the unsupervised autoencoder. Change the value of the variables dataset, batch_size, time_length and iterations according to the requirements. Comment prep.normalize_input() [line 128] and prep.create_input() [line 131], if data preprocessing is not required. Sync the value of time_length in the preprocessing file with the value in the training file.
  2. supervised_training.py Command: python3 supervised_training.py Code for training the supervised autoencoder. Same as above, for changes and comments.

Testing:

  1. testing.py Command: python3 testing.py > output.txt here, output file will store the reconstruction error per frame. Ensure that the values of time_length, batch_size and dataset are same as the values during training. Output is only the reconstruction error of all the frames of all the test videos. In order to execute the script evaluation and creation of summaries, manually create files corresponding to the video names and copy-paste the reconstruction errors. Check the path where the individual files (check in the summary creation script)
  2. testing_Grad_CAM.py Command: python3 testing_Grad_CAM.py > output.txt Ensure the values of the variable as mentioned earlier. Output is a tuple <reconstruction error, frame number> for each video of the test set. For summary creation manually create files corresponding to the video names and copy-paste the reconstruction errors. Check the path where the individual files (check in the summary creation script)

Anomaly Summarization:

  1. thresholding.py Command: python3 thresholding.py <video_number> Uses the thresholding mechanism, described in the paper, for creating summaries. For creating summaries of the entire test dataset, use the helper script generate_result.sh. For executing the shell script, write the following command: ./generate_result.sh As output, we get a summary video corresponding to the video number
  2. sorting_selection.py Command: python3 sorting_selection.py <video_number> Uses the sorting and selection mechanism for creating summaries. For creating summaries of the entire test dataset, use the helper script generate_result.sh. For executing the shell script, write the following command: ./generate_result.sh As output, we get a summary video corresponding to the video number
  3. sliding_window.py Command: python3 sliding_window.py <video_number> Slides a window across the duration of the entire video and picks that segment which has the maximum reconstruction error. For creating summaries of the entire test dataset, use the helper script generate_result.sh. For executing the shell script, write the following command: ./generate_result.sh As output, we get a summary video corresponding to the video number
  4. generate_result.sh Command: ./generate_result.sh <number_of_videos> Runs a loop from 1 to n, where n is the total number of testing videos. So, ensure that the videos are numbered from 1 to n. Depending on the kind of summary required, change the file name inside this file. 2 commands have been commented for illustration sake.

Helper Scripts:

  1. parse_json.py Command: python3 parse_json.py Takes a json file, which has been output after rating segments using oTranscribe tool and produces groundtruth/rating files in text format. Change the input file in the program manually and place the input json file in the same directory. Update the value of factor, fps and dataset as required.
  2. extract_frames.sh Command: ./extract_frames.sh Change the value of the variable corresponding to dataset, training_video_folder and testing_video_folder according to your requirements. Contains extraction of all the frames and saving them as both .jpg and .png. Comment relevant portion, if not required.
  3. Few other scripts have also been published. These are not directly related to this work, but can be useful for future endeavours. extract_darknet_features.sh extracts darknet features of each frame of all the videos. Output is a file for each frame. These files are combined to form a single feature file for each video using the program proprocess_after_darknet.py preprocess_after_darknet.py combines the files containing the extracted features and creates a combined feature file for each video.