Human Detection using Hardware Oriented GMM-MRCoHOG

Abstract

The purpose of this study is to reduce calculation cost of Gaussian Mixture Model - Mul- tiResolution Co-occurrence Histogram of Oriented Gradients (GMM-MRCoHOG) feature for hardware implementation. This feature shows high accuracy for human detection but has problems at hardware implementation: it requires inverse trigonometric function for gradient computation and large amount of memory resources for storing parameters. So, We proposed two methods for solving each problem. First, use conditional branch instead of inverse trigonometric function. Second, use fuzzy inference for saving memory resources. The experimental results show that these methods maintain high accuracy with reduce in calculation cost and memory resources.