An Efficient Method of HOG Feature Extraction Using Selective Histogram Bin and PCA Feature Reduction

Histogram of Oriented Gradient (HOG) is a popular image feature for human detection. It presents high detection accuracy and therefore has been widely used in vision-based surveillance and pedestrian detection systems. However, the main drawback of this feature is that it has a large feature size. The extraction algorithm is also computationally intensive and requires long processing time. In this paper, a time-efficient HOG-based feature extraction method is proposed. The method uses selective number of histogram bins to perform feature extraction on different regions in the image. Higher number of histogram bin which can capture more detailed information is performed on the regions of the image which may belong to part of a human figure, while lower number of histogram bin is used on the rest of the image. To further reduce the feature size, Principal Component Analysis (PCA) is used to rank the features and remove some unimportant features. The performance of the proposed method was evaluated using INRIA human dataset on a linear Support Vector Machine (SVM) classifier. The results showed the processing speed of the proposed method is 2.6 times faster than the original HOG and 7 times faster than the LBP method while providing comparable detection performance.