Locally normalized Histogram of Oriented Gradient (HOG) algorithm originally proposed by Dalal & Triggs presents excellent results for pedestrian detection. However, as the demand of accuracy and speed in real-time application increase, the detection speed and robustness of this method is becoming insufficient. Over the years, improvements have been proposed by different researchers in order to meet the requirement of the robustness and processing speed. This includes the improvement in the ways HOG feature is extracted, combination of HOG feature with other image features and using part based detection method. This paper reviews the current advancement in HOG features for human detection.