Each day in 2019, around 2,000 petabytes of video data are generated by security cameras around the world, up from 500 petabytes per day in 2015 which is equivalent to 75 million users streaming an hour’s HDTV simultaneously (IHS Markit, Jan 26, 2016). This dramatic growth is driving the demand for more efficient methods of analysing video data.
Commonly used deep neural networks such as Convolutional Neural Networks (CNN) have demonstrated outstanding performance for video analysis tasks, but the training process for CNNs can take days to months to complete.
While methods based on hand crafted feature descriptors require shorter training and processing times, the high- dimensionality of the extracted features demands vast storage capacities and computational resources.
New analytics from the University of Warwick are able to address these needs and can be used for classification, action recognition, object recognition, video surveillance, monitoring and abnormal event detection.