While Histograms of Gradients (HOG) has been proven as an excellent feature set for human detection, its intricate computation does not readily lend itself to be realized into high performance and economical hardware. This paper proposes several methods to simplify complicated computations of HOG, so that HOG feature extraction can be speeded up, which is essential in real-time applications, and become more appealing. Considering that a successful detection does not depend on the precision of individual elements in the descriptor, our intention is to exploit this to simplify the HOG arithmetic. In other words, with no discernible impact on detection, our aim is to accelerate the computations of HOG by reducing the accuracy of the arithmetic with a controlled upper bound. The proposed methods have been utilized on widely known test dataset, ETHZ, and the results show that no detection performance is suffered.