Abstract :
Driven by the need of efficient video classification techniques to detect abnormal activities among Surveillance videos, we propose an method for classifying videos to either be anomalous(containing some abnormality) or normal(contains no abnormality). Most existing video abnormality detection techniques are based on the presumption that anything that deviates from normal pattern is considered to be abnormal. However in realistic scenarios where we consider different surveillance videos at different locations or different angles of view, it is extremely difficult to define normality of an event. Also, what may be normal event in one environment may not be normal in other. To make the problem more realistic, we plan to work on dataset which contains videos from different scenarios such as road, cafe, etc. Since the above assumption won't hold for such diverse scenarios, we train our model on both normal and anomalous videos. Also, our method only uses the video level labels for training the model.
LINK to Report : https://drive.google.com/file/d/1juQtzemcOzf6C08xMddQiMdhT-zLiqzY/view?usp=sharing
LINK to Video Results :
DATASET:
You can also download dataset in parts through following link
https://www.dropbox.com/sh/75v5ehq4cdg5g5g/AABvnJSwZI7zXb8_myBA0CLHa?dl=0
Below you can find Training/Testing Code for our anomaly Detection project as part of COMPSCI 682:
The implementation is tested using:
Keras version 1.1.0
Theano 1.0.2
Python 3
Ubuntu 16.04
We used C3D-v1.0 (https://github.com/facebook/C3D) with default settings as a feature extractor.
Training_AnomalyDetecor_public.py is to Train Anomaly Detection Model
Testing_Anomaly_Detector_public.py is to test trained Anomaly Detection Model
Save_C3DFeatures_32Segments is to save already computed C3D features for the whole video into 32 segment features.
weights_L1L2.mat: It contains the pre-trained weights for the model ‘model.json’.
Demo_GUI: We have provided a simple GUI which can be used to see results of our approach on sample videos.
SampleVideos: This folder contains C3D features (32 segments) for sample videos. It order to see testing results for the features in this folder, please copy the corrosponding videos in the same folder.
Plot_All_ROC: This code can be use to plot the ROC results reported in the paper. The data to plot ROCs of methods discussed in the paper can be found in folder Paper_Results.
Temporal_Anomaly_Annotation.txt contains ground truth annotations of the testing dataset.
Anomaly_Train.txt contains the video names for training anomaly detector