CUDA implementation of ViBe background subtraction algorithm on jetson TX1/TX2 modules

With the emergence of smart IP cameras capable of end-node processing, a substantial subset of the algorithms aimed at Surveillance and Smart Cities are required to run on NVIDIA Jetson TX1 / TX2 modules. Background subtraction is a pre-processing block that is frequently used by video analytic algorithms. In this work, a CUDA implementation of the ViBe (Visual Background Extraction) background subtraction algorithm on NVIDIA Jetson TX1 / TX2 modules will be presented. The optimizations that are required for real-time execution on Jetson TX1 / TX2 modules with 720p video will be described and several strategies that can help enhance performance of video analytic algorithms that leverage the GPU of these modules will be presented.