Poisson Mixture Model for High Speed and Low-Power Background Subtraction

Abstract

Background subtraction is an important computer vision scheme for segmenting moving objects from a video. The imminent IoT era imposes severe constraints on camera systems in terms of area, speed, power, and accuracy. This chapter presents a new background subtraction scheme which is based on Poisson noise model, where we use a Poisson mixture model (PMM) for modeling dynamic backgrounds. Furthermore, we have used a method for sequential estimation of parameters of the PMM which does not allow the parameters to fluctuate with random noise. Resultantly, our method achieves a reduction of 17.76% and 8.16% in percent of wrong classification compared to the hardware implementations of Gaussian Mixture Models and Pixel-based Adaptive Segmented (PBAS) methods, respectively. The proposed scheme is also very fast, low power, and requires very little logic resources. The memory bandwidth requirement of the proposed scheme is 6.66%, 41.36%, and 90.48% lower compared to state-of-the-art FPGA implementation of GMM, ViBe, and PBAS algorithms, respectively. The proposed background subtraction scheme achieves a significant speed up compared to FPGA implementations of GMM (by 42.68%), ViBe (by 115%), and PBAS (by 142.2%) schemes. Similarly, the energy consumption of the proposed method is 76.39% and 99.9% less compared to GMM and PBAS, respectively. In summary, the advantages of the proposed method in accuracy, speed, and energy consumption combined together make it especially suitable for embedded applications.