Fast background modeling using GMM on GPU

Although Gaussian Mixture Model (GMM) for background modeling can give better results, it is still too slow to be applied to real-time systems. This paper describes two different parallel implementations on GPU to accelerate the GMM for background subtracting (BGS). One is basic GPU implementation using constant memory, which just assigns a GPU thread for each pixel and doesn´t do much optimization. The other is asynchronous GPU implementation, which employs different optimizations techniques, such as pinned memory, memory coalescing, and asynchronous execution. Both of the two implementations are benefit from the computational capacity of CUDA cores on GPUs. The experimental results show our GPU implementation for background subtraction can greatly save the modeling time without obvious degradation of accuracy.