Single Core SIMD Parallelization of GMM Background Subtraction Algorithm for Vehicles Detection

Gaussian Mixture Model background subtraction (GMM) method is nowadays used in many moving object detection applications. This pixel wise approach is performed by applying the same arithmetic operations on independent data elements. Thus, the vector execution of these operations is an efficient acceleration method; especially because most of modern SIMD architectures implement long registers in their processors. In this paper, we propose an efficient data vectorization of GMM on two different Intel architectures using the SSE2 vector instructions. This is done by parallelizing the hotspots that consume big processing time in the algorithm, as well as using suitable compilation environment. The performance evaluation for varying datasets using only one core showed that a speedup of 1,8 is achieved. This is satisfactory for this algorithm where many portions are sequential, and where most of computation is done in single-precision floating-point. Hence, this has limited the number of elements that we could pack in the core registers.