Event Detection in Video Surveillance System

Abstract

Surveillance videos are ubiquitous in the safety critical places. Detecting any type of abnormality in the place using intelligent surveillance system is an active area of research. Traditionally, handcrafted features like HoG (Histogram of Gradients) are used to detect objects from video. Recently, deep learning techniques have outperformed the state of the art object detection techniques. Convolutional Neural Networks (CNN) are used to solve many of the computer vision problems. CNNs learn the features required for detecting object or event from video frames unlike the traditional systems where features are obtained by a user written program. In this report, we have done the comparison of the HoG features and learned features for object and behavior detection by implementing two separate machine learning models, Support Vector Machines (SVM) and K-means classifier on custom data-set. Here, performance of the detection system with both features was analyzed. We have also done event detection using CNN and LSTM with the help of GPU and high speed processor. Event detection using CNN, uses only spatial data for classification of video with the help of convolutional neural network. Event detection using LSTM, learns features using pre-trained modal (Inception-V3) and classify video on both spatial and temporal data with the help of LSTM network.