Benchmark Dataset Creation for Intelligent Surveillance System under Dense Environment (BDISSDE)

Surveillance has always been of utmost importance but most of the time we see a person looking at CCTV footage for security or no one watching them at all. Present systems have helped to identify the criminals after the crime has been committed, but have not yet been able to take an immediate action. Up until now the researches have either manually created a benchmark dataset of classified actions or used half manual half automation approach. Despite advancement in image processing the challenges have hardly been touched in some of the scenarios. Like low visibility due to mist/fog or night, live multiple moving objects, rainy condition or even snow. The authors propose a fully automated mechanism to create a digitized benchmark dataset which would help to identify a moving object under any condition and its action as well. Preprocessing of each frame is carried out using a model which helps to extract the features and create a classified dataset which could be used to identify the action. Bag of features, SIFT, Optical flow and saliency features have been extracted and stored to prepare a benchmark dataset. The dataset has been tested for robustness using benchmark algorithms like SVM, CNN, KNN. KTH, 20 bn datasets have been used for the benchmark action videos, other videos, under different conditions, have been shot by the author. It is found that BDISSDE performs exceptionally well under various algorithms.