High performance video processing in cloud data centres

Abstract

Mobile phones and affordable cameras are generating large amounts of video data. This data holds information regarding several activities and incidents. Video analytics systems have been introduced to extract valuable information from this data. However, most of these systems are expensive, require human supervision and are time consuming. The probability of extracting inaccurate information is also high due to human involvement. We have addressed these challenges by proposing a cloud based high performance video analytics platform. This platform attempts to minimize human intervention, reduce computation time and enables the processing of a large number of video streams. It achieves high performance by optimizing the occupancy of GPU resources in cloud and minimizing the data transfer by concurrently processing a large number of video streams. The proposed video processing platform is evaluated in three stages. The first evaluation was performed at the cloud level in order to evaluate the scalability of the platform. This evaluation includes fetching and distributing video streams and efficiently utilizing available resources within the cloud. The second valuation was performed at the individual cloud nodes. This evaluation includes measuring the occupancy level, effect of data transfer and the extent of concurrency achieved at each node. The third evaluation was performed at the frame level in order to determine the performance of object recognition algorithms. To measure this, compute intensive tasks of the Local Binary Pattern (LBP) algorithm have been ported on to the GPU resources. The platform proved to be very scalable with high throughput and performance when tested on a large number of video streams with increasing number of nodes.