Implementation of intrusion detection system in CUDA for real-time multi-node streaming

Abstract

A common surveillance activity is to track important people, or people exhibiting suspicious behavior, as they move from one camera surveillance area to another. The reduction in video hardware cost has made it more feasible for large scale camera deployment. However, the increased scale of camera deployment creates difficulties for humans to track people through the monitored space and to recognize important events as they happen in timely manner without human intervention. In this paper we share the implementation of the multi node video analytics specifically focusing on intrusion detection. The system uses general purpose graphical processing unit (GPGPU) to offload the video analytics processing. The architecture of the GPGPU requires the algorithm to be coded in Compute Unified Device Architecture (CUDA) which involves algorithm parallelization adopting both micro and macro parallelization to ensure the performance gain in processing speed on per frame basis by 7 times. In addition, we have managed to deploy 35 camera streams on single GPU card running at 20 frames per second which results in scalability factor of 1.75 times vs. a server class PC. Indeed, we have also managed to maintain the video analytics accuracy at 100% for given test dataset, in this implementation of the system.